Structured Computer Organization: United States Edition [5 ed.] 0131485210, 9780131485211

Citation preview

STRUCTURED COMPUTER ORGANIZATION

Other bestselling titles by Andrew S. Tanenbaum Computer Networks, 4th edition This widely-read classic, now in its fourth edition, provides the ideal introduction to today’s and tomorrow’s networks. It explains in detail how modern networks are structured. Starting with the physical layer and working up to the application layer, the book covers a vast number of important topics, including wireless communication, fiber optics, data link protocols, Ethernet, routing algorithms, network performance, security, DNS, electronic mail, USENET news, the World Wide Web, and multimedia. The book has especially thorough coverage of TCP/IP and the Internet.

Operating Systems: Design and Implementation, 2nd edition This popular text on operating systems, co-authored with Albert S. Woodhull, is the only book covering both the principles of operating systems and their application to a real system. All the traditional operating systems topics are covered in detail. In addition, the principles are carefully illustrated with MINIX, a free POSIX-based UNIX-like operating system for personal computers. Each book contains a free CD-ROM containing the complete MINIX system, including all the source code. The source code is listed in an appendix to the book and explained in detail in the text.

Modern Operating Systems, 2nd edition This comprehensive text covers the principles of modern operating systems in detail and illustrates them with numerous real-world examples. After an introductory chapter, the next five chapters deal with the basic concepts: processes and threads, deadlocks, memory management, input/output, and file systems. The next six chapters deal with more advanced material, including multimedia systems, multiple processor systems, security. Finally, two detailed case studies are given: UNIX/Linux and Windows 2000.

Distributed Systems: Principles and Paradigms This new book, co-authored with Maarten van Steen, covers both the principles and paradigms of modern distributed systems. In the first part, it covers the principles of communication, processes, naming, synchronization, consistency and replication, fault tolerance, and security in detail. Then in the second part, it goes into different paradigms used to build distributed systems, including object-based systems, distributed file systems, document-based systems, and coordinationbased systems.

STRUCTURED COMPUTER ORGANIZATION FIFTH EDITION

ANDREW S. TANENBAUM Vrije Universiteit Amsterdam, The Netherlands

UPPER SADDLE RIVER, NEW JERSEY 07458

Library of Congress Cataloging-in-Publication Data Tanenbaum, Andrew S. Structured computer organization / Andrew S. Tanenbaum. —5th edition p. cm. Includes bibliographical references and index. ISBN 0-13-148521-0 1. Computer programming. 2. Computer organization. I. Title. QA76.6.T38 2005 005.1--dc22 2005043004

Vice President and Editorial Director, ECS: Marcia Horton Senior Acquisitions Editor: Tracy Dunkelberger Editorial Assistant: Michael Giacobbe Executive Managing Editor: Vince O'Brien Managing Editor: Camille Trentacoste Production Editor: Irwin Zucker Director of Creative Services: Paul Belfanti Art Director: Heather Scott Cover Illustrator: Maraska Artistic Concept and Design Don Martinetti, DM Graphics, Inc. Cover Concept: Andrew S. Tanenbaum Composition and interior design: Andrew S. Tanenbaum Managing Editor, AV Management and Production: Patricia Burns Art Editor: Gregory Dulles Manufacturing Buyer: Lisa McDowell Marketing Manager: Pamela Hersperger Marketing Assistant: Barrie Reinhold © 2006, 1999, 1990, 1984, 1976 by Pearson Education, Inc. Pearson Prentice Hall Pearson Education, Inc. Upper Saddle River, NJ 07458 All rights reserved. No part of this book may be reproduced in any form or by any means, without permission in writing from the publisher. Pearson Prentice Hall® is a trademark of Pearson Education, Inc. The author and publisher of this book have used their best efforts in preparing this book. These efforts include the development, research, and testing of the theories and programs to determine their effectiveness. The author and publisher make no warranty of any kind, expressed or implied, with regard to these programs or the documentation contained in this book. The author and publisher shall not be liable in any event for incidental or consequential damages in connection with, or arising out of, the furnishing, performance, or use of these programs. Printed in the United States of America 10 9 8 7 6 5 4 3 2 1

ISBN

0-13-148521-0

Pearson Education Ltd., London Pearson Education Australia Pty. Ltd., Sydney Pearson Education Singapore, Pte. Ltd. Pearson Education North Asia Ltd., Hong Kong Pearson Education Canada, Inc., Toronto Pearson Educación de Mexico, S.A. de C.V. Pearson Education-Japan, Tokyo Pearson Education Malaysia, Pte. Ltd. Pearson Education, Inc., Upper Saddle River, New Jersey

To Suzanne, Barbara, Marvin, and the memory of Sweetie π and Bram

This page intentionally left blank

CONTENTS xviii

PREFACE

1

1

INTRODUCTION 1.1

STRUCTURED COMPUTER ORGANIZATION 2 1.1.1 Languages, Levels, and Virtual Machines 2 1.1.2 Contemporary Multilevel Machines 5 1.1.3 Evolution of Multilevel Machines 8

1.2

MILESTONES IN COMPUTER ARCHITECTURE 13 1.2.1 The Zeroth Generation—Mechanical Computers (1642–1945) 14 1.2.2 The First Generation—Vacuum Tubes (1945–1955) 16 1.2.3 The Second Generation—Transistors (1955–1965) 19 1.2.4 The Third Generation—Integrated Circuits (1965–1980) 22 1.2.5 The Fourth Generation—Very Large Scale Integration (1980-?) 23 1.2.6 The Fifth Generation—Invisible Computers 26

1.3

THE COMPUTER ZOO 27 1.3.1 Technological and Economic Forces 1.3.2 The Computer Spectrum 29 1.3.3 Disposable Computers 29 1.3.4 Microcontrollers 31 1.3.5 Game Computers 33 1.3.6 Personal Computers 34 vii

27

viii

CONTENTS

1.3.7 Servers 34 1.3.8 Collections of Workstations 1.3.9 Mainframes 36

2

34

1.4

EXAMPLE COMPUTER FAMILIES 37 1.4.1 Introduction to the Pentium 4 37 1.4.2 Introduction to the UltraSPARC III 42 1.4.3 Introduction to the 8051 44

1.5

METRIC UNITS 46

1.6

OUTLINE OF THIS BOOK 47

COMPUTER SYSTEMS ORGANIZATION 2.1

2.2

2.3

PROCESSORS 51 2.1.1 CPU Organization 52 2.1.2 Instruction Execution 54 2.1.3 RISC versus CISC 58 2.1.4 Design Principles for Modern Computers 2.1.5 Instruction-Level Parallelism 61 2.1.6 Processor-Level Parallelism 65 PRIMARY MEMORY 69 2.2.1 Bits 69 2.2.2 Memory Addresses 70 2.2.3 Byte Ordering 71 2.2.4 Error-Correcting Codes 73 2.2.5 Cache Memory 77 2.2.6 Memory Packaging and Types SECONDARY MEMORY 81 2.3.1 Memory Hierarchies 81 2.3.2 Magnetic Disks 82 2.3.3 Floppy Disks 86 2.3.4 IDE Disks 86 2.3.5 SCSI Disks 88 2.3.6 RAID 89 2.3.7 CD-ROMs 93

80

59

51

ix

CONTENTS

2.3.8 CD-Recordables 97 2.3.9 CD-Rewritables 99 2.3.10 DVD 99 2.3.11 Blu-Ray 102 2.4

2.5

3

INPUT/OUTPUT 102 2.4.1 Buses 102 2.4.2 Terminals 105 2.4.3 Mice 110 2.4.4 Printers 112 2.4.5 Telecommunications Equipment 2.4.6 Digital Cameras 125 2.4.7 Character Codes 127

117

SUMMARY 131

135

THE DIGITAL LOGIC LEVEL 3.1

GATES AND BOOLEAN ALGEBRA 135 3.1.1 Gates 136 3.1.2 Boolean Algebra 138 3.1.3 Implementation of Boolean Functions 3.1.4 Circuit Equivalence 141

3.2

BASIC DIGITAL LOGIC CIRCUITS 146 3.2.1 Integrated Circuits 146 3.2.2 Combinational Circuits 147 3.2.3 Arithmetic Circuits 152 3.2.4 Clocks 157

3.3

MEMORY 159 3.3.1 Latches 159 3.3.2 Flip-Flops 161 3.3.3 Registers 163 3.3.4 Memory Organization 164 3.3.5 Memory Chips 168 3.3.6 RAMs and ROMs 171

140

x

CONTENTS

3.4

CPU CHIPS AND BUSES 173 3.4.1 CPU Chips 174 3.4.2 Computer Buses 176 3.4.3 Bus Width 178 3.4.4 Bus Clocking 180 3.4.5 Bus Arbitration 184 3.4.6 Bus Operations 187

3.5

EXAMPLE CPU CHIPS 189 3.5.1 The Pentium 4 189 3.5.2 The UltraSPARC III 196 3.5.3 The 8051 200

3.6

EXAMPLE BUSES 202 3.6.1 The ISA Bus 203 3.6.2 The PCI Bus 204 3.6.3 PCI Express 212 3.6.4 The Universal Serial Bus

3.7

3.8

4

INTERFACING 221 3.7.1 I/O Chips 221 3.7.2 Address Decoding

217

222

SUMMARY 225

THE MICROARCHITECTURE LEVEL 4.1

AN EXAMPLE MICROARCHITECTURE 231 4.1.1 The Data Path 232 4.1.2 Microinstructions 239 4.1.3 Microinstruction Control: The Mic-1 241

4.2

AN EXAMPLE ISA: IJVM 246 4.2.1 Stacks 246 4.2.2 The IJVM Memory Model 248 4.2.3 The IJVM Instruction Set 250 4.2.4 Compiling Java to IJVM 254

231

xi

CONTENTS

4.3

5

AN EXAMPLE IMPLEMENTATION 255 4.3.1 Microinstructions and Notation 255 4.3.2 Implementation of IJVM Using the Mic-1

260

4.4

DESIGN OF THE MICROARCHITECTURE LEVEL 271 4.4.1 Speed versus Cost 271 4.4.2 Reducing the Execution Path Length 273 4.4.3 A Design with Prefetching: The Mic-2 281 4.4.4 A Pipelined Design: The Mic-3 281 4.4.5 A Seven-Stage Pipeline: The Mic-4 288

4.5

IMPROVING PERFORMANCE 292 4.5.1 Cache Memory 293 4.5.2 Branch Prediction 299 4.5.3 Out-of-Order Execution and Register Renaming 4.5.4 Speculative Execution 309

304

4.6

EXAMPLES OF THE MICROARCHITECTURE LEVEL 311 4.6.1 The Microarchitecture of the Pentium 4 CPU 312 4.6.2 The Microarchitecture of the UltraSPARC-III Cu CPU 317 4.6.3 The Microarchitecture of the 8051 CPU 323

4.7

COMPARISON OF THE PENTIUM, ULTRASPARC, AND 8051 325

4.8

SUMMARY 326

THE INSTRUCTION SET ARCHITECTURE LEVEL 331 5.1

OVERVIEW OF THE ISA LEVEL 333 5.1.1 Properties of the ISA Level 333 5.1.2 Memory Models 335 5.1.3 Registers 337 5.1.4 Instructions 339 5.1.5 Overview of the Pentium 4 ISA Level 339 5.1.6 Overview of the UltraSPARC III ISA Level 341 5.1.7 Overview of the 8051 ISA Level 345

xii

CONTENTS

5.2

DATA TYPES 348 5.2.1 Numeric Data Types 348 5.2.2 Nonnumeric Data Types 349 5.2.3 Data Types on the Pentium 4 350 5.2.4 Data Types on the UltraSPARC III 350 5.2.5 Data Types on the 8051 351

5.3

INSTRUCTION FORMATS 351 5.3.1 Design Criteria for Instruction Formats 352 5.3.2 Expanding Opcodes 354 5.3.3 The Pentium 4 Instruction Formats 357 5.3.4 The UltraSPARC III Instruction Formats 358 5.3.5 The 8051 Instruction Formats 359

5.4

5.5

ADDRESSING 360 5.4.1 Addressing Modes 360 5.4.2 Immediate Addressing 361 5.4.3 Direct Addressing 361 5.4.4 Register Addressing 361 5.4.5 Register Indirect Addressing 362 5.4.6 Indexed Addressing 363 5.4.7 Based-Indexed Addressing 365 5.4.8 Stack Addressing 365 5.4.9 Addressing Modes for Branch Instructions 369 5.4.10 Orthogonality of Opcodes and Addressing Modes 369 5.4.11 The Pentium 4 Addressing Modes 371 5.4.12 The UltraSPARC III Addressing Modes 373 5.4.13 The 8051 Addressing Modes 373 5.4.14 Discussion of Addressing Modes 374 INSTRUCTION TYPES 375 5.5.1 Data Movement Instructions 375 5.5.2 Dyadic Operations 376 5.5.3 Monadic Operations 377 5.5.4 Comparisons and Conditional Branches 5.5.5 Procedure Call Instructions 381 5.5.6 Loop Control 382 5.5.7 Input/Output 383 5.5.8 The Pentium 4 Instructions 386 5.5.9 The UltraSPARC III Instructions 389 5.5.10 The 8051 Instructions 392 5.5.11 Comparison of Instruction Sets 392

379

xiii

CONTENTS

5.6

FLOW OF CONTROL 395 5.6.1 Sequential Flow of Control and Branches 5.6.2 Procedures 396 5.6.3 Coroutines 401 5.6.5 Traps 404 5.6.5 Interrupts 404

5.7

A DETAILED EXAMPLE: THE TOWERS OF HANOI 408 5.7.1 The Towers of Hanoi in Pentium 4 Assembly Language 409 5.7.2 The Towers of Hanoi in UltraSPARC III Assembly Language 409

5.8

THE IA-64 ARCHITECTURE AND THE ITANIUM 2 411 5.8.1 The Problem with the Pentium 4 413 5.8.2 The IA-64 Model: Explicitly Parallel Instruction Computing 5.8.3 Reducing Memory References 415 5.8.4 Instruction Scheduling 416 5.8.5 Reducing Conditional Branches: Predication 418 5.8.6 Speculative Loads 420

5.9

6

395

414

SUMMARY 421

THE OPERATING SYSTEM MACHINE LEVEL 6.1

VIRTUAL MEMORY 428 6.1.1 Paging 429 6.1.2 Implementation of Paging 431 6.1.3 Demand Paging and the Working Set Model 433 6.1.4 Page Replacement Policy 436 6.1.5 Page Size and Fragmentation 438 6.1.6 Segmentation 439 6.1.7 Implementation of Segmentation 442 6.1.8 Virtual Memory on the Pentium 4 445 6.1.9 Virtual Memory on the UltraSPARC III 450 6.1.10 Virtual Memory and Caching 452

6.2

VIRTUAL I/O INSTRUCTIONS 453 6.2.1 Files 454 6.2.2 Implementation of Virtual I/O Instructions 6.2.3 Directory Management Instructions 459

455

427

xiv

7

CONTENTS

6.3

VIRTUAL INSTRUCTIONS FOR PARALLEL PROCESSING 460 6.3.1 Process Creation 461 6.3.2 Race Conditions 462 6.3.3 Process Synchronization Using Semaphores 466

6.4

EXAMPLE OPERATING SYSTEMS 470 6.4.1 Introduction 470 6.4.2 Examples of Virtual Memory 479 6.4.3 Examples of Virtual I/O 482 6.4.4 Examples of Process Management 493

6.5

SUMMARY 500

THE ASSEMBLY LANGUAGE LEVEL 7.1

INTRODUCTION TO ASSEMBLY LANGUAGE 508 7.1.1 What Is an Assembly Language? 508 7.1.2 Why Use Assembly Language? 509 7.1.3 Format of an Assembly Language Statement 512 7.1.4 Pseudoinstructions 515

7.2

MACROS 517 7.2.1 Macro Definition, Call, and Expansion 518 7.2.2 Macros with Parameters 520 7.2.3 Advanced Features 521 7.2.4 Implementation of a Macro Facility in an Assembler

7.3

THE ASSEMBLY PROCESS 522 7.3.1 Two-Pass Assemblers 522 7.3.2 Pass One 523 7.3.3 Pass Two 527 7.3.4 The Symbol Table 529

7.4

LINKING AND LOADING 530 7.4.1 Tasks Performed by the Linker 532 7.4.2 Structure of an Object Module 535 7.4.3 Binding Time and Dynamic Relocation 7.4.4 Dynamic Linking 539

7.5

SUMMARY 543

536

507

521

xv

CONTENTS

8

PARALLEL COMPUTER ARCHITECTURES 8.1

ON-CHIP PARALELLISM 548 8.1.1 Instruction-Level Parallelism 549 8.1.2 On-Chip Multithreading 556 8.1.3 Single-Chip Multiprocessors 562

8.2

COPROCESSORS 567 8.2.1 Network Processors 568 8.2.2 Media Processors 576 8.2.3 Cryptoprocessors 581

8.3

SHARED-MEMORY MULTIPROCESSORS 582 8.3.1 Multiprocessors vs. Multicomputers 582 8.3.2 Memory Semantics 590 8.3.3 UMA Symmetric Multiprocessor Architectures 8.3.4 NUMA Multiprocessors 602 8.3.5 COMA Multiprocessors 611

8.4

9

MESSAGE-PASSING MULTICOMPUTERS 612 8.4.1 Interconnection Networks 614 8.4.2 MPPs—Massively Parallel Processors 617 8.4.3 Cluster Computing 627 8.4.4 Communication Software for Multicomputers 8.4.5 Scheduling 635 8.4.6 Application-Level Shared Memory 636 8.4.7 Performance 643

8.5

GRID COMPUTING 649

8.6

SUMMARY 651

READING LIST AND BIBLIOGRAPHY 9.1

SUGGESTIONS FOR FURTHER READING 655 9,1,1 Introduction and General Works 655 9.1.2 Computer Systems Organization 657 9.1.3 The Digital Logic Level 658

547

594

632

655

xvi

CONTENTS

9.1.4 The Microarchitecture Level 659 9.1.5 The Instruction Set Architecture Level 659 9.1.6 The Operating System Machine Level 660 9.1.7 The Assembly Language Level 661 9.1.8 Parallel Computer Architectures 661 9.1.9 Binary and Floating-Point Numbers 663 9.1.10 Assembly Language Programming 664 9.2

A

B

C

ALPHABETICAL BIBLIOGRAPHY 664

679

BINARY NUMBERS A.1

FINITE-PRECISION NUMBERS 679

A.2

RADIX NUMBER SYSTEMS 681

A.3

CONVERSION FROM ONE RADIX TO ANOTHER 683

A.4

NEGATIVE BINARY NUMBERS 685

A.5

BINARY ARITHMETIC 688

691

FLOATING-POINT NUMBERS B.1

PRINCIPLES OF FLOATING POINT 692

B.2

IEEE FLOATING-POINT STANDARD 754 694

ASSEMBLY LANGUAGE PROGRAMMING C.1

OVERVIEW 702 C.1.1 Assembly Language 702 C.1.2 A Small Assembly Language Program

703

701

xvii

CONTENTS

C.2

THE 8088 PROCESSOR 704 C.2.1 The Processor Cycle 705 C.2.2 The General Registers 705 C.2.3 Pointer Registers 708

C.3

MEMORY AND ADDRESSING 709 C.3.1 Memory Organization and Segments C.3.2 Addressing 711

C.4

THE 8088 INSTRUCTION SET 715 C.4.1 Move, Copy and Arithmetic 715 C.4.2 Logical, Bit and Shift Operations 718 C.4.3 Loop and Repetitive String Operations 718 C.4.4 Jump and Call Instructions 719 C.4.5 Subroutine Calls 721 C.4.6 System Calls and System Subroutines 723 C.4.7 Final Remarks on the Instruction Set 725

C.5

THE ASSEMBLER 725 C.5.1 Introduction 726 C.5.2 The ACK-Based Tutorial Assembler as88 727 C.5.3 Some Differences with Other 8088 Assemblers 730

C.6

THE TRACER 732 C.6.1 Tracer Commands

C.7

GETTING STARTED 735

C.8

EXAMPLES 736 C.8.1 Hello World Example 736 C.8.2 General Registers Example 740 C.8.3 Call Command and Pointer Registers 742 C.8.4 Debugging an Array Print Program 744 C.8.5 String Manipulation and String Instructions C.8.6 Dispatch Tables 750 C.8.7 Buffered and Random File Access 752

INDEX

709

734

748

757

PREFACE

The first four editions of this book were based on the idea that a computer can be regarded as a hierarchy of levels, each one performing some well-defined function. This fundamental concept is as valid today as it was when the first edition came out, so it has been retained as the basis for the fifth edition. As in the first four editions, the digital logic level, the microarchitecture level, the instruction set architecture level, the operating system machine level, and the assembly language level are all discussed in detail. Although the basic structure has been maintained, this fifth edition does contain many changes, both small and large, that bring it up to date in the rapidly changing computer industry. For example, the example machines used have been brought up to date. The current examples are the Intel Pentium 4, the Sun UltraSPARC III, and the Intel 8051. The Pentium 4 is an example of a popular CPU used on desktop machines. The UltraSPARC III is an example of a popular server, widely used in medium and large mutiprocessors. However, the 8051 may come as a surprise to some people. It is a venerable chip that has been around for decades. However, with the enormous growth of embedded systems, it has finally come into its own. With computers running everything from clock radios to microwave ovens, interest in embedded systems is surging, and the 8051 is a widely-used choice due to its extremely low cost (pennies), the wealth of software and peripherals for it, and the large number of 8051 programmers available. Over the years, many professors teaching from the course have repeatedly asked for material on assembly language programming. With the fifth edition, that material is now available in Appendix C and on the accompanying CD-ROM. xviii

PREFACE

xix

The assembly language chosen is the 8088 since it is a stripped down version of the enormously popular Pentium. I could have used the UltraSPARC or the MIPS or some other CPU almost no one has ever heard of, but as a motivational tool, the 8088 is a better choice since large numbers of students have a Pentium at home and the Pentium is capable of running 8088 programs. However, since debugging assembly code is very difficult, I have provided a set of tools for learning assembly language programming, including an 8088 assembler, a simulator, and a tracer. These tools are provided for Windows UNIX, and Linux. The tools are on the CD-ROM and also on the book’s Website (see below). The book has become longer over the years. Such an expansion is inevitable as a subject develops and there is more known about it. As a result, when the book is used for a course, it may not always be possible to finish the book in a single course (e.g., in a trimester system). A possible approach would be to do all of Chaps. 1, 2, and 3, the first part of Chap. 4 (up through and including Sec. 4.4), and Chap. 5 as a bare minimum. The remaining time could be filled with the rest of Chap. 4, and parts of Chaps. 6, 7, and 8, depending on the interest of the instructor. A chapter-by-chapter rundown of the major changes since the fourth edition follows. Chapter 1 still contains an historical overview of computer architecture, pointing out how we got where we are now and what the milestones were along the way. The enlarged spectrum of computers that exist is now discussed, and our three major examples (Pentium 4, UltraSPARC III, and 8051) are introduced. In Chapter 2, the material on input/output devices has been updated, emphasizing the technology of modern devices, including digital cameras, DSL, and Internet over cable. Chapter 3 has undergone some revision and now treats computer buses and modern I/O chips. The three new examples are described here at the chip level. New material has been added about the PCI Express bus, which is expected to replace the PCI bus shortly. Chapter 4 has always been a popular chapter for explaining how a computer really works, so most of it is unchanged since the fourth edition. However, there are new sections discussing the microarchitecture level of Pentium 4, the UltraSPARC III, and the 8051. Chapters 5, 6, and 7 have been updated using the new examples, but are otherwise relatively unchanged. Chapter 6 uses Windows XP rather than Windows NT as an example, but at the level of discussion here, the changes are minimal. In contrast, Chapter 8 has been heavily modified to reflect all the new activity in parallel computers of all forms. It covers five different classes of parallel systems, from on-chip parallelism (instruction-level parallelism, on-chip multithreading, and single-chip multiprocessors), through coprocessors, shared-memory systems, and clusters, and ends up with a brief discussion of grids. Numerous new examples are discussed here, from the TriMedia CPU, to the BlueGene/L, Red Storm and Google clusters.

xx

PREFACE

The references in Chap. 9 have been updated heavily. Computer organization is a dynamic field. Over half the references in this 5th edition are to books and papers written after the 4th edition of this book was published. Appendices A and B are unchanged since last time, but Appendix C on assembly language programming is completely new. It is a hands-on, how-to guide to assembly language programming using the tools provided on the CD-ROM and the Website. Appendix C was written by Dr. Evert Wattel of the Vrije Universiteit, Amsterdam. Dr. Wattel has had many years of experience teaching students using these tools. My thanks to him for writing this appendix. In addition to the assembly language tools, the Website also contains a graphical simulator to be used in conjunction with Chap. 4. This simulator was written by Prof. Richard Salter of Oberlin College. It can be used by students to help grasp the principles discussed in this chapter. My thanks to him for providing this software. In addition, the figures used in the book and PowerPoint sheets for instructors are also available on the Website. The URL is http:/www.prenhall.com/tanenbaum From there, click on the Companion Website for this book and select the page you are looking for from the menu. Instructors using this book for a university course can obtain a manual with the solutions to the problems by contacting their Pearson Education representative. A number of people have read (parts of) the manuscript and provided useful suggestions or have been helpful in other ways. In particular, I would like to thank Nikitas Alexandridis, Shekar Borkar, Herbert Bos, Scott Cannon, Doug Carmean, Alan Charlesworth, Eric Cota-Robles, Michael Fetterman, Quinn Jacobson, Thilo Kielmann, Iffat Kazi, Saul Levy, Ahmed Louri, Abhijit Pandya, Krist Petersen, Mark Russinovich, Ronald Schroeder, and Saim Ural for their help, for which I am most grateful. Thank you. I would also like to thank Jim Goodman for his contributions to this book, especially to Chaps, 4 and 5. The idea of using the Java Virtual Machine was his and the book is better for it. Finally, I would like to thank Suzanne once more for her love and patience. It never ends, not even after 15 books. Barbara and Marvin are always a joy and now know what professors do for a living. The Royal Netherlands Academy of Arts and Sciences granted me a much-coveted Academy Professorship in 2004, freeing me from some of the less attractive aspects of academia (such as endless boring committee meetings), for which I am eternally grateful. Andrew S. Tanenbaum

1 INTRODUCTION

A digital computer is a machine that can solve problems for people by carrying out instructions given to it. A sequence of instructions describing how to perform a certain task is called a program. The electronic circuits of each computer can recognize and directly execute a limited set of simple instructions into which all its programs must be converted before they can be executed. These basic instructions are rarely much more complicated than Add two numbers. Check a number to see if it is zero. Copy a piece of data from one part of the computer’s memory to another. Together, a computer’s primitive instructions form a language in which people can communicate with the computer. Such a language is called a machine language. The people designing a new computer must decide what instructions to include in its machine language. Usually, they try to make the primitive instructions as simple as possible, consistent with the computer’s intended use and performance requirements, in order to reduce the complexity and cost of the electronics needed. Because most machine languages are so simple, it is difficult and tedious for people to use them. This simple observation has, over the course of time, led to a way of structuring computers as a series of abstractions, each abstraction building on the one 1

2

INTRODUCTION

CHAP. 1

below it. In this way, the complexity can be mastered and computer systems can be designed in a systematic, organized way. We call this approach structured computer organization and have named the book after it. In the next section we will describe what we mean by this term. After that we will look at some historical developments, the state-of-the-art, and some important examples.

1.1 STRUCTURED COMPUTER ORGANIZATION As mentioned above, there is a large gap between what is convenient for people and what is convenient for computers. People want to do X, but computers can only do Y. This leads to a problem. The goal of this book is to explain how this problem can be solved.

1.1.1 Languages, Levels, and Virtual Machines The problem can be attacked in two ways: both involve designing a new set of instructions that is more convenient for people to use than the set of built-in machine instructions. Taken together, these new instructions also form a language, which we will call L1, just as the built-in machine instructions form a language, which we will call L0. The two approaches differ in the way programs written in L1 are executed by the computer, which, after all, can only execute programs written in its machine language, L0. One method of executing a program written in L1 is first to replace each instruction in it by an equivalent sequence of instructions in L0. The resulting program consists entirely of L0 instructions. The computer then executes the new L0 program instead of the old L1 program. This technique is called translation. The other technique is to write a program in L0 that takes programs in L1 as input data and carries them out by examining each instruction in turn and executing the equivalent sequence of L0 instructions directly. This technique does not require first generating a new program in L0. It is called interpretation and the program that carries it out is called an interpreter. Translation and interpretation are similar. In both methods, the computer carries out instructions in L1 by executing equivalent sequences of instructions in L0. The difference is that, in translation, the entire L1 program is first converted to an L0 program, the L1 program is thrown away, and then the new L0 program is loaded into the computer’s memory and executed. During execution, the newly generated L0 program is running and in control of the computer. In interpretation, after each L1 instruction is examined and decoded, it is carried out immediately. No translated program is generated. Here, the interpreter is in control of the computer. To it, the L1 program is just data. Both methods, and increasingly, a combination of the two, are widely used.

SEC. 1.1

STRUCTURED COMPUTER ORGANIZATION

3

Rather than thinking in terms of translation or interpretation, it is often simpler to imagine the existence of a hypothetical computer or virtual machine whose machine language is L1. Let us call this virtual machine M1 (and let us call the virtual machine corresponding to L0, M0). If such a machine could be constructed cheaply enough, there would be no need for having language L0 or a machine that executed programs in L0 at all. People could simply write their programs in L1 and have the computer execute them directly. Even if the virtual machine whose language is L1 is too expensive or complicated to construct out of electronic circuits, people can still write programs for it. These programs can either be interpreted or translated by a program written in L0 that itself can be directly executed by the existing computer. In other words, people can write programs for virtual machines, just as though they really existed. To make translation or interpretation practical, the languages L0 and L1 must not be ‘‘too’’ different. This constraint often means that L1, although better than L0, will still be far from ideal for most applications. This result is perhaps discouraging in light of the original purpose for creating L1— relieving the programmer of the burden of having to express algorithms in a language more suited to machines than people. However, the situation is not hopeless. The obvious approach is to invent still another set of instructions that is more people-oriented and less machine-oriented than L1. This third set also forms a language, which we will call L2 (and with virtual machine M2). People can write programs in L2 just as though a virtual machine with L2 as its machine language really existed. Such programs can either be translated to L1 or executed by an interpreter written in L1. The invention of a whole series of languages, each one more convenient than its predecessors, can go on indefinitely until a suitable one is finally achieved. Each language uses its predecessor as a basis, so we may view a computer using this technique as a series of layers or levels, one on top of another, as shown in Fig. 1-1. The bottommost language or level is the simplest and the topmost language or level is the most sophisticated. There is an important relation between a language and a virtual machine. Each machine has a machine language, consisting of all the instructions that the machine can execute. In effect, a machine defines a language. Similarly, a language defines a machine—namely, the machine that can execute all programs written in the language. Of course, the machine defined by a certain language may be enormously complicated and expensive to construct directly out of electronic circuits but we can imagine it nevertheless. A machine with C or C++ or Java as its machine language would be complex indeed but could easily be built using today’s technology. There is a good reason, however, for not building such a computer: it would not be cost effective compared to other techniques. Merely being doable is not good enough: a practical design must be cost effective as well. In a certain sense, a computer with n levels can be regarded as n different virtual machines, each with a different machine language. We will use the terms

4

INTRODUCTION

Virtual machine Mn, with machine language Ln

Level 3

Virtual machine M3, with machine language L3

…

Level n

Level 2

Level 1

Level 0

CHAP. 1

Programs in Ln are either interpreted by an interpreter running on a lower machine, or are translated to the machine language of a lower machine

Virtual machine M2, with machine language L2

Programs in L2 are either interpreted by interpreters running on M1 or M0, or are translated to L1 or L0

Virtual machine M1, with machine language L1

Programs in L1 are either interpreted by an interpreter running on M0, or are translated to L0

Actual computer M0, with machine language L0

Programs in L0 can be directly executed by the electronic circuits

Figure 1-1. A multilevel machine.

‘‘level’’ and ‘‘virtual machine’’ interchangeably. Only programs written in language L0 can be directly carried out by the electronic circuits, without the need for intervening translation or interpretation. Programs written in L1, L2, ..., Ln must either be interpreted by an interpreter running on a lower level or translated to another language corresponding to a lower level. A person who writes programs for the level n virtual machine need not be aware of the underlying interpreters and translators. The machine structure ensures that these programs will somehow be executed. It is of no real interest whether they are carried out step by step by an interpreter which, in turn, is also carried out by another interpreter, or whether they are carried out by the electronic circuits directly. The same result appears in both cases: the programs are executed. Most programmers using an n-level machine are interested only in the top level, the one least resembling the machine language at the very bottom. However, people interested in understanding how a computer really works must study all the levels. People who design new computers or new levels (i.e., new virtual machines) must also be familiar with levels other than the top one. The concepts and techniques of constructing machines as a series of levels and the details of the levels themselves form the main subject of this book.

SEC. 1.1

LANGUAGES, LEVELS, AND VIRTUAL MACHINES

5

1.1.2 Contemporary Multilevel Machines Most modern computers consist of two or more levels. Machines with as many as six levels exist, as shown in Fig. 1-2. Level 0, at the bottom, is the machine’s true hardware. Its circuits carry out the machine-language programs of level 1. For the sake of completeness, we should mention the existence of yet another level below our level 0. This level, not shown in Fig. 1-2 because it falls within the realm of electrical engineering (and is thus outside the scope of this book), is called the device level. At this level, the designer sees individual transistors, which are the lowest-level primitives for computer designers. If one asks how transistors work inside, that gets us into solid-state physics. Level 5

Problem-oriented language level Translation (compiler)

Level 4

Assembly language level Translation (assembler)

Level 3

Operating system machine level Partial interpretation (operating system)

Level 2

Instruction set architecture level Interpretation (microprogram) or direct execution

Level 1

Microarchitecture level Hardware

Level 0

Digital logic level

Figure 1-2. A six-level computer. The support method for each level is indicated below it (along with the name of the supporting program).

At the lowest level that we will study, the digital logic level, the interesting objects are called gates. Although built from analog components, such as transistors, gates can be accurately modeled as digital devices. Each gate has one or more digital inputs (signals representing 0 or 1) and computes as output some simple function of these inputs, such as AND or OR. Each gate is built up of at most a handful of transistors. A small number of gates can be combined to form a 1-bit memory, which can store a 0 or a 1. The 1-bit memories can be combined in groups of (for example) 16, 32, or 64 to form registers. Each register can hold a

6

INTRODUCTION

CHAP. 1

single binary number up to some maximum. Gates can also be combined to form the main computing engine itself. We will examine gates and the digital logic level in detail in Chap. 3. The next level up is the microarchitecture level. At this level we see a collection of (typically) 8 to 32 registers that form a local memory and a circuit called an ALU (Arithmetic Logic Unit), which is capable of performing simple arithmetic operations. The registers are connected to the ALU to form a data path, over which the data flow. The basic operation of the data path consists of selecting one or two registers, having the ALU operate on them (for example, adding them together), and storing the result stored back in some register. On some machines the operation of the data path is controlled by a program called a microprogram. On other machines the data path is controlled directly by hardware. In the first three editions of this book, we called this level the ‘‘microprogramming level,’’ because in the past it was nearly always a sofware interpreter. Since the data path is now often (partially) controlled directly by hardware, we changed the name in the previous edition to reflect this. On machines with software control of the data path, the microprogram is an interpreter for the instructions at level 2. It fetches, examines, and executes instructions one by one, using the data path to do so. For example, for an ADD instruction, the instruction would be fetched, its operands located and brought into registers, the sum computed by the ALU, and finally the result routed back to the place it belongs. On a machine with hardwired control of the data path, similar steps would take place, but without an explicit stored program to control the interpretation of the level 2 instructions. At level 2 we have a level that we will call the Instruction Set Architecture level or (ISA level). Every computer manufacturer publishes a manual for each of the computers it sells, entitled ‘‘Machine Language Reference Manual’’ or ‘‘Principles of Operation of the Western Wombat Model 100X Computer’’ or something similar. These manuals are really about the ISA level, not the underlying levels. When they describe the machine’s instruction set, they are in fact describing the instructions carried out interpretively by the microprogram or hardware execution circuits. If a computer manufacturer provides two interpreters for one of its machines, interpreting two different ISA levels, it will need to provide two ‘‘machine language’’ reference manuals, one for each interpreter. The next level is usually a hybrid level. Most of the instructions in its language are also in the ISA level. (There is no reason why an instruction appearing at one level cannot be present at other levels as well.) In addition, there is a set of new instructions, a different memory organization, the ability to run two or more programs concurrently, and various other features. More variation exists between level 3 designs than between those at either level 1 or level 2. The new facilities added at level 3 are carried out by an interpreter running at level 2, which, historically, has been called an operating system. Those level 3 instructions that are identical to level 2’s are carried out directly by the micro-

SEC. 1.1

LANGUAGES, LEVELS, AND VIRTUAL MACHINES

7

program (or hardwired control), not by the operating system. In other words, some of the level 3 instructions are interpreted by the operating system and some are interpreted directly by the microprogram. This is what we mean by ‘‘hybrid’’ level. Throughout this book we will call this level the operating system machine level. There is a fundamental break between levels 3 and 4. The lowest three levels are not designed for use by the average garden-variety programmer. Instead they are intended primarily for running the interpreters and translators needed to support the higher levels. These interpreters and translators are written by people called systems programmers who specialize in designing and implementing new virtual machines. Levels 4 and above are intended for the applications programmer with a problem to solve. Another change occurring at level 4 is the method by which the higher levels are supported. Levels 2 and 3 are always interpreted. Levels 4, 5, and above are usually, although not always, supported by translation. Yet another difference between levels 1, 2, and 3, on the one hand, and levels 4, 5, and higher, on the other, is the nature of the language provided. The machine languages of levels 1, 2, and 3 are numeric. Programs in them consist of long series of numbers, which are fine for machines but bad for people. Starting at level 4, the languages contain words and abbreviations meaningful to people. Level 4, the assembly language level, is really a symbolic form for one of the underlying languages. This level provides a method for people to write programs for levels 1, 2, and 3 in a form that is not as unpleasant as the virtual machine languages themselves. Programs in assembly language are first translated to level 1, 2, or 3 language and then interpreted by the appropriate virtual or actual machine. The program that performs the translation is called an assembler. Level 5 usually consists of languages designed to be used by applications programmers with problems to solve. Such languages are often called high-level languages. Literally hundreds exist. A few of the better known ones are C, C++, Java, LISP, and Prolog. Programs written in these languages are generally translated to level 3 or level 4 by translators known as compilers, although occasionally they are interpreted instead. Programs in Java, for example, are usually first translated to a an ISA-like language called Java byte code, which is then interpreted. In some cases, level 5 consists of an interpreter for a specific application domain, such as symbolic mathematics. It provides data and operations for solving problems in this domain in terms that people knowledgeable in that domain can understand easily. In summary, the key thing to remember is that computers are designed as a series of levels, each one built on its predecessors. Each level represents a distinct abstraction, with different objects and operations present. By designing and analyzing computers in this fashion, we are temporarily able to suppress irrelevant detail and thus reduce a complex subject to something easier to understand.

8

INTRODUCTION

CHAP. 1

The set of data types, operations, and features of each level is called its architecture. The architecture deals with those aspects that are visible to the user of that level. Features that the programmer sees, such as how much memory is available, are part of the architecture. Implementation aspects, such as what kind of technology is used to implement the memory, are not part of the architecture. The study of how to design those parts of a computer system that are visible to the programmers is called computer architecture. In common practice, however, computer architecture and computer organization mean essentially the same thing.

1.1.3 Evolution of Multilevel Machines To provide some perspective on multilevel machines, we will briefly examine their historical development, showing how the number and nature of the levels has evolved over the years. Programs written in a computer’s true machine language (level 1) can be directly executed by the computer’s electronic circuits (level 0), without any intervening interpreters or translators. These electronic circuits, along with the memory and input/output devices, form the computer’s hardware. Hardware consists of tangible objects—integrated circuits, printed circuit boards, cables, power supplies, memories, and printers—rather than abstract ideas, algorithms, or instructions. Software, in contrast, consists of algorithms (detailed instructions telling how to do something) and their computer representations—namely, programs. Programs can be stored on hard disk, floppy disk, CD-ROM, or other media, but the essence of software is the set of instructions that makes up the programs, not the physical media on which they are recorded. In the very first computers, the boundary between hardware and software was crystal clear. Over time, however, it has blurred considerably, primarily due to the addition, removal, and merging of levels as computers have evolved. Nowadays, it is often hard to tell them apart (Vahid, 2003). In fact, a central theme of this book is Hardware and software are logically equivalent. Any operation performed by software can also be built directly into the hardware, preferably after it is sufficiently well understood. As Karen Panetta Lentz put it: ‘‘Hardware is just petrified software.’’ Of course, the reverse is also true: any instruction executed by the hardware can also be simulated in software. The decision to put certain functions in hardware and others in software is based on such factors as cost, speed, reliability, and frequency of expected changes. There are few hard-and-fast rules to the effect that X must go into the hardware and Y must be programmed explicitly. These decisions change with trends in technology economics, demand, and computer usage.

SEC. 1.1

LANGUAGES, LEVELS, AND VIRTUAL MACHINES

9

The Invention of Microprogramming The first digital computers, back in the 1940s, had only two levels: the ISA level, in which all the programming was done, and the digital logic level, which executed these programs. The digital logic level’s circuits were complicated, difficult to understand and build, and unreliable. In 1951, Maurice Wilkes, a researcher at the University of Cambridge, suggested designing a three-level computer in order to drastically simplify the hardware (Wilkes, 1951). This machine was to have a built-in, unchangeable interpreter (the microprogram), whose function was to execute ISA-level programs by interpretation. Because the hardware would now only have to execute microprograms, which have a limited instruction set, instead of ISA-level programs, which have a much larger instruction set, fewer electronic circuits would be needed. Because electronic circuits were then made from vacuum tubes, such a simplification promised to reduce tube count and hence enhance reliability (i.e., the number of crashes per day). A few of these three-level machines were constructed during the 1950s. More were constructed during the 1960s. By 1970 the idea of having the ISA level be interpreted by a microprogram, instead of directly by the electronics, was dominant. All the major machines of the day used it. The Invention of the Operating System In these early years, most computers were ‘‘open shop,’’ which meant that the programmer had to operate the machine personally. Next to each machine was a sign-up sheet. A programmer wanting to run a program signed up for a block of time, say Wednesday morning 3 to 5 A.M. (many programmers liked to work when it was quiet in the machine room). When the time arrived, the programmer headed for the machine room with a deck of 80-column punched cards (an early input medium) in one hand and a sharpened pencil in the other. Upon arriving in the computer room, he or she gently nudged the previous programmer toward the door and took over the computer. If the programmer wanted to run a FORTRAN program, the following steps were necessary: 1. He† went over to the cabinet where the program library was kept, took out the big green deck labeled FORTRAN compiler, put it in the card reader, and pushed the START button. 2. He put his FORTRAN program in the card reader and pushed the CONTINUE button. The program was read in.

""""""""""""""""""""""""""""""""""""""""""""""""

† ‘‘He’’ should be read as ‘‘he or she’’ throughout this book.

10

INTRODUCTION

CHAP. 1

3. When the computer stopped, he read his FORTRAN program in a second time. Although some compilers required only one pass over the input, many required two or more. For each pass, a large card deck had to be read in. 4. Finally, the translation neared completion. The programmer often became nervous near the end because if the compiler found an error in the program, he had to correct it and start the entire process all over again. If there were no errors, the compiler punched out the translated machine language program on cards. 5. The programmer then put the machine language program in the card reader along with the subroutine library deck and read them both in. 6. The program began executing. More often than not it did not work and unexpectedly stopped in the middle. Generally, the programmer fiddled with the console switches and looked at the console lights for a little while. If lucky, he figured out the problem, corrected the error, and went back to the cabinet containing the big green FORTRAN compiler to start all over again. If less fortunate, he made a printout of the contents of memory, called a core dump, and took it home to study. This procedure, with minor variations, was normal at many computer centers for years. It forced the programmers to learn how to operate the machine and to know what to do when it broke down, which was often. The machine was frequently idle while people were carrying cards around the room or scratching their heads trying to find out why their programs were not working properly. Around 1960 people tried to reduce the amount of wasted time by automating the operator’s job. A program called an operating system was kept in the computer at all times. The programmer provided certain control cards along with the program that were read and carried out by the operating system. Figure 1-3 shows a sample job for one of the first widespread operating systems, FMS (FORTRAN Monitor System), on the IBM 709. The operating system read the ∗JOB card and used the information on it for accounting purposes. (The asterisk was used to identify control cards, so they would not be confused with program and data cards.) Later, it read the ∗FORTRAN card, which was an instruction to load the FORTRAN compiler from a magnetic tape. The compiler then read in and compiled the FORTRAN program. When the compiler finished, it returned control back to the operating system, which then read the ∗DATA card. This was an instruction to execute the translated program, using the cards following the ∗DATA card as the data. Although the operating system was designed to automate the operator’s job (hence the name), it was also the first step in the development of a new virtual machine. The ∗FORTRAN card could be viewed as a virtual ‘‘compile program’’

SEC. 1.1

LANGUAGES, LEVELS, AND VIRTUAL MACHINES

11

*JOB, 5494, BARBARA *XEQ *FORTRAN

FORTRAN program

*DATA

Data cards

*END Figure 1-3. A sample job for the FMS operating system.

instruction. Similarly, the *DATA card could be regarded as a virtual ‘‘execute program’’ instruction. A level with only two instructions was not much of a level but it was a start in that direction. In subsequent years, operating systems became more and more sophisticated. New instructions, facilities, and features were added to the ISA level until it began to take on the appearance of a new level. Some of this new level’s instructions were identical to the ISA-level instructions, but others, particularly input/output instructions, were completely different. The new instructions were often known as operating system macros or supervisor calls. The usual term now is system call. Operating systems developed in other ways as well. The early ones read card decks and printed output on the line printer. This organization was known as a batch system. Usually, there was a wait of several hours between the time a program was submitted and the time the results were ready. Developing software was difficult under those circumstances. In the early 1960s researchers at Dartmouth College, M.I.T., and elsewhere developed operating systems that allowed (multiple) programmers to communicate directly with the computer. In these systems, remote terminals were connected to the central computer via telephone lines. The computer was shared among many users. A programmer could type in a program and get the results typed back almost immediately, in the office, in a garage at home, or wherever the terminal was located. These systems were called timesharing systems.

12

INTRODUCTION

CHAP. 1

Our interest in operating systems is in those parts that interpret the instructions and features present in level 3 and not present in the ISA level rather than in the timesharing aspects. Although we will not emphasize it, you should keep in mind that operating systems do more than just interpret features added to the ISA level. The Migration of Functionality to Microcode Once microprogramming had become common (by 1970), designers realized that they could add new instructions by just extending the microprogram. In other words, they could add ‘‘hardware’’ (new machine instructions) by programming. This revelation led to a virtual explosion in machine instruction sets, as designers competed with one another to produce bigger and better instruction sets. Many of these instructions were not essential in the sense that their effect could be easily achieved by existing instructions, but often they were slightly faster than a sequence of existing instructions. For example, many machines had an instruction INC (INCrement) that added one to a number. Since these machines also had a general ADD instruction, having a special instruction to add 1 (or to add 720, for that matter) was not necessary. However, the INC was usually a little faster than the ADD, so it got thrown in. For the same reason, many other instructions were added to the microprogram. These often included 1. Instructions for integer multiplication and division. 2. Floating-point arithmetic instructions. 3. Instructions for calling and returning from procedures. 4. Instructions for speeding up looping. 5. Instructions for handling character strings. Furthermore, once machine designers saw how easy it was to add new instructions, they began looking around for other features to add to their microprograms. A few examples of these additions include 1. Features to speed up computations involving arrays (indexing and indirect addressing). 2. Features to permit programs to be moved in memory after they have started running (relocation facilities). 3. Interrupt systems that signal the computer as soon as an input or output operation is completed. 4. The ability to suspend one program and start another in a small number of instructions (process switching).

SEC. 1.1

LANGUAGES, LEVELS, AND VIRTUAL MACHINES

13

5. Special instructions for processing audio, image, and multimedia files. Numerous other features and facilities have been added over the years as well, usually for speeding up some particular activity. The Elimination of Microprogramming Microprograms grew fat during the golden years of microprogramming (1960s and 1970s). They also tended to get slower and slower as they acquired more bulk. Finally, some researchers realized that by eliminating the microprogram, vastly reducing the instruction set, and having the remaining instructions be directly executed (i.e., hardware control of the data path), machines could be speeded up. In a certain sense, computer design had come full circle, back to the way it was before Wilkes invented microprogramming in the first place. But the wheel is still turning. Java programs are generally executed by compiling them to an intermediate language (Java byte code), and then interpreting the Java byte code. The point of this discussion is to show that the boundary between hardware and software is arbitrary and constantly changing. Today’s software may be tomorrow’s hardware, and vice versa. Furthermore, the boundaries between the various levels are also fluid. From the programmer’s point of view, how an instruction is actually implemented is unimportant (except perhaps for its speed). A person programming at the ISA level can use its multiply instruction as though it were a hardware instruction without having to worry about it, or even be aware of whether it really is a hardware instruction. One person’s hardware is another person’s software. We will come back to all these topics later in this book.

1.2 MILESTONES IN COMPUTER ARCHITECTURE Hundreds of different kinds of computers have been designed and built during the evolution of the modern digital computer. Most have been long forgotten, but a few have had a significant impact on modern ideas. In this section we will give a brief sketch of some of the key historical developments in order to get a better understanding of how we got where we are now. Needless to say, this section only touches on the highlights and leaves many stones unturned. Figure 1-4 lists some of the milestone machines to be discussed in this section. Slater (1987) is a good place to look for additional historical material on the people who founded the computer age. For short biographies and beautiful color photographs by Louis Fabian Bachrach of some of the key people who founded the computer age, see Morgan’s coffee-table book (1997).

14

INTRODUCTION

CHAP. 1

##################################################################################### ! Year ! ! ! ! Name Made by Comments ##################################################################################### ! ! ! ! ! ! 1834 ! Analytical Engine ! Babbage ! First attempt to build a digital computer ! ##################################################################################### ! 1936 ! Z1 ! Zuse ! First working relay calculating machine ! ##################################################################################### ! ! ! ! ! ! 1943 ! COLOSSUS ! British gov’t ! First electronic computer ! ##################################################################################### ! ! ! ! ! 1944 Mark I Aiken First American general-purpose computer ##################################################################################### ! ! ! ! ! ! 1946 ! ENIAC I ! Eckert/Mauchley ! Modern computer history starts here ! ##################################################################################### ! ! ! ! ! ! 1949 ! EDSAC ! Wilkes ! First stored-program computer ! ##################################################################################### ! 1951 ! Whirlwind I ! M.I.T. ! First real-time computer ! ##################################################################################### ! ! ! ! ! ! 1952 ! IAS ! Von Neumann ! Most current machines use this design ! ##################################################################################### ! 1960 ! PDP-1 ! DEC ! First minicomputer (50 sold) ! ##################################################################################### ! ! ! ! ! ! 1961 ! 1401 ! IBM ! Enormously popular small business machine ! ##################################################################################### ! ! ! ! ! ##################################################################################### ! 1962 ! 7094 ! IBM ! Dominated scientific computing in the early 1960s ! ! 1963 ! B5000 ! Burroughs ! First machine designed for a high-level language ! ##################################################################################### ! ! ! ! ! ! 1964 ! 360 ! IBM ! First product line designed as a family ! ##################################################################################### ! 1964 ! 6600 ! CDC ! First scientific supercomputer ! ##################################################################################### ! ! ! ! ! ! 1965 ! PDP-8 ! DEC ! First mass-market minicomputer (50,000 sold) ! ##################################################################################### ! ! ! ! ! ##################################################################################### ! 1970 ! PDP-11 ! DEC ! Dominated minicomputers in the 1970s ! ! 1974 ! 8080 ! Intel ! First general-purpose 8-bit computer on a chip ! ##################################################################################### ! ! ! ! ! ##################################################################################### ! 1974 ! CRAY-1 ! Cray ! First vector supercomputer ! ! 1978 ! VAX ! DEC ! First 32-bit superminicomputer ! ##################################################################################### ! ! ! ! ! ! 1981 ! IBM PC ! IBM ! Started the modern personal computer era ! ##################################################################################### ! 1981 ! Osborne-1 ! Osborne ! First portable computer ! ##################################################################################### ! ! ! ! ! ! 1983 ! Lisa ! Apple ! First personal computer with a GUI ! ##################################################################################### ! ! ! ! ! ##################################################################################### ! 1985 ! 386 ! Intel ! First 32-bit ancestor of the Pentium line ! ! 1985 ! MIPS ! MIPS ! First commercial RISC machine ! ##################################################################################### ! ! ! ! ! ! 1987 ! SPARC ! Sun ! First SPARC-based RISC workstation ! ##################################################################################### ! 1990 ! RS6000 ! IBM ! First superscalar machine ! ##################################################################################### ! ! ! ! ! ! 1992 ! Alpha ! DEC ! First 64-bit personal computer ! ##################################################################################### ! 1993 ! Newton ! Apple ! First palmtop computer ! ! ##################################################################################### ! ! ! ! Figure 1-4. Some milestones in the development of the modern digital computer.

1.2.1 The Zeroth Generation—Mechanical Computers (1642–1945) The first person to build a working calculating machine was the French scientist Blaise Pascal (1623–1662), in whose honor the programming language Pascal is named. This device, built in 1642, when Pascal was only 19, was designed to help his father, a tax collector for the French government. It was entirely mechanical, using gears, and powered by a hand-operated crank.

SEC. 1.2

MILESTONES IN COMPUTER ARCHITECTURE

15

Pascal’s machine could only do addition and subtraction operations, but thirty years later the great German mathematician Baron Gottfried Wilhelm von Leibniz (1646–1716) built another mechanical machine that could multiply and divide as well. In effect, Leibniz had built the equivalent of a four-function pocket calculator three centuries ago. Nothing much happened for 150 years until a professor of mathematics at the University of Cambridge, Charles Babbage (1792–1871), the inventor of the speedometer, designed and built his difference engine. This mechanical device, which like Pascal’s could only add and subtract, was designed to compute tables of numbers useful for naval navigation. The entire construction of the machine was designed to run a single algorithm, the method of finite differences using polynomials. The most interesting feature of the difference engine was its output method: it punched its results into a copper engraver’s plate with a steel die, thus foreshadowing later write-once media such as punched cards and CD-ROMs. Although the difference engine worked reasonably well, Babbage quickly got bored with a machine that could run only one algorithm. He began to spend increasingly large amounts of his time and family fortune (not to mention 17,000 pounds of the government’s money) on the design and construction of a successor called the analytical engine. The analytical engine had four components: the store (memory), the mill (computation unit), the input section (punched card reader), and the output section (punched and printed output). The store consisted of 1000 words of 50 decimal digits, each used to hold variables and results. The mill could accept operands from the store, then add, subtract, multiply, or divide them, and finally return the result to the store. Like the difference engine, it was entirely mechanical. The great advance of the analytical engine was that it was general purpose. It read instructions from punched cards and carried them out. Some instructions commanded the machine to fetch two numbers from the store, bring them to the mill, be operated on (e.g., added), and have the result sent back to the store. Other instructions could test a number and conditionally branch depending on whether it was positive or negative. By punching a different program on the input cards, it was possible to have the analytical engine perform different computations, something not true of the difference engine. Since the analytical engine was programmable in a simple assembly language, it needed software. To produce this software, Babbage hired a young woman named Ada Augusta Lovelace, who was the daughter of the famed British poet, Lord Byron. Ada Lovelace was thus the world’s first computer programmer. The ® programming language Ada is named in her honor. Unfortunately, like many modern designers, Babbage never quite got the hardware debugged. The problem was that he needed thousands upon thousands of cogs and wheels and gears produced to a degree of precision that nineteenthcentury technology was unable to provide. Nevertheless, his ideas were far ahead of his time, and even today most modern computers have a structure very similar

16

INTRODUCTION

CHAP. 1

to the analytical engine, so it is certainly fair to say that Babbage was the (grand)father of the modern digital computer. The next major development occurred in the late 1930s, when a German engineering student named Konrad Zuse built a series of automatic calculating machines using electromagnetic relays. He was unable to get government funding after the war began because government bureaucrats expected to win the war so quickly that the new machine would not be ready until after it was over. Zuse was unaware of Babbage’s work, and his machines were destroyed by the Allied bombing of Berlin in 1944, so his work did not have any influence on subsequent machines. Still, he was one of the pioneers of the field. Slightly later, in the United States, two people also designed calculators, John Atanasoff at Iowa State College and George Stibbitz at Bell Labs. Atanasoff’s machine was amazingly advanced for its time. It used binary arithmetic and had capacitors for memory, which were periodically refreshed to keep the charge from leaking out, a process he called ‘‘jogging the memory.’’ Modern dynamic memory (DRAM) chips work the same way. Unfortunately the machine never really became operational. In a way, Atanasoff was like Babbage: a visionary who was ultimately defeated by the inadequate hardware technology of his time. Stibbitz’ computer, although more primitive than Atanasoff’s, actually worked. Stibbitz gave a public demonstration of it at a conference at Dartmouth College in 1940. One of the people in the audience was John Mauchley, an unknown professor of physics at the University of Pennsylvania. The computing world would hear more about Prof. Mauchley later. While Zuse, Stibbitz, and Atanasoff were designing automatic calculators, a young man named Howard Aiken was grinding out tedious numerical calculations by hand as part of his Ph.D. research at Harvard. After graduating, Aiken recognized the importance of being able to do calculations by machine. He went to the library, discovered Babbage’s work, and decided to build out of relays the general-purpose computer that Babbage had failed to build out of toothed wheels. Aiken’s first machine, the Mark I, was completed at Harvard in 1944. It had 72 words of 23 decimal digits each, and had an instruction time of 6 sec. Input and output used punched paper tape. By the time Aiken had completed its successor, the Mark II, relay computers were obsolete. The electronic era had begun.

1.2.2 The First Generation—Vacuum Tubes (1945–1955) The stimulus for the electronic computer was World War II. During the early part of the war, German submarines were wreaking havoc on British ships. Commands were sent from the German admirals in Berlin to the submarines by radio, which the British could, and did, intercept. The problem was that these messages were encoded using a device called the ENIGMA, whose forerunner was designed by amateur inventor and former U.S. president, Thomas Jefferson.

SEC. 1.2

MILESTONES IN COMPUTER ARCHITECTURE

17

Early in the war, British intelligence managed to acquire an ENIGMA machine from Polish Intelligence, which had stolen it from the Germans. However, to break a coded message, a huge amount of computation was needed, and it was needed very soon after the message was intercepted to be of any use. To decode these messages, the British government set up a top secret laboratory that built an electronic computer called the COLOSSUS. The famous British mathematician Alan Turing helped design this machine. The COLOSSUS was operational in 1943, but since the British government kept virtually every aspect of the project classified as a military secret for 30 years, the COLOSSUS line was basically a dead end. It is only worth noting because it was the world’s first electronic digital computer. In addition to destroying Zuse’s machines and stimulating the construction of the COLOSSUS, the war also affected computing in the United States. The army needed range tables for aiming its heavy artillery. It produced these tables by hiring hundreds of women to crank them out using hand calculators (women were thought to be more accurate than men). Nevertheless, the process was time consuming and errors often crept in. John Mauchley, who knew of Atanasoff’s work as well as Stibbitz’, was aware that the army was interested in mechanical calculators. Like many computer scientists after him, he put together a grant proposal asking the army for funding to build an electronic computer. The proposal was accepted in 1943, and Mauchley and his graduate student, J. Presper Eckert, proceeded to build an electronic computer, which they called the ENIAC (Electronic Numerical Integrator And Computer). It consisted of 18,000 vacuum tubes and 1500 relays. The ENIAC weighed 30 tons and consumed 140 kilowatts of power. Architecturally, the machine had 20 registers, each capable of holding a 10-digit decimal number. (A decimal register is very small memory that can hold one number up to some maximum number of decimal digits, somewhat like the odometer that keeps track of how far a car has traveled in its lifetime.) The ENIAC was programmed by setting up 6000 multiposition switches and connecting a multitude of sockets with a veritable forest of jumper cables. The machine was not finished until 1946, when it was too late to be of any use for its original purpose. However, since the war was over, Mauchley and Eckert were allowed to organize a summer school to describe their work to their scientific colleagues. That summer school was the beginning of an explosion of interest in building large digital computers. After that historic summer school, many other researchers set out to build electronic computers. The first one operational was the EDSAC (1949), built at the University of Cambridge by Maurice Wilkes. Others included the JOHNIAC at the Rand Corporation, the ILLIAC at the University of Illinois, the MANIAC at Los Alamos Laboratory, and the WEIZAC at the Weizmann Institute in Israel. Eckert and Mauchley soon began working on a successor, the EDVAC (Electronic Discrete Variable Automatic Computer). However, that project was

18

INTRODUCTION

CHAP. 1

fatally wounded when they left the University of Pennsylvania to form a startup company, the Eckert-Mauchley Computer Corporation, in Philadelphia (Silicon Valley had not yet been invented). After a series of mergers, this company became the modern Unisys Corporation. As a legal aside, Eckert and Mauchley filed for a patent claiming they invented the digital computer. In retrospect, this would not be a bad patent to own. After years of litigation, the courts decided that the Eckert-Mauchley patent was invalid and that John Atanasoff invented the digital computer, even though he never patented it. While Eckert and Mauchley were working on the EDVAC, one of the people involved in the ENIAC project, John von Neumann, went to Princeton’s Institute of Advanced Studies to build his own version of the EDVAC, the IAS machine. Von Neumann was a genius in the same league as Leonardo Da Vinci. He spoke many languages, was an expert in the physical sciences and mathematics, and had total recall of everything he ever heard, saw, or read. He was able to quote from memory the verbatim text of books he had read years earlier. At the time he became interested in computers, he was already the most eminent mathematician in the world. One of the things that was soon apparent to him was that programming computers with huge numbers of switches and cables was slow, tedious, and inflexible. He came to realize that the program could be represented in digital form in the computer’s memory, along with the data. He also saw that the clumsy serial decimal arithmetic used by the ENIAC, with each digit represented by 10 vacuum tubes (1 on and 9 off) could be replaced by using parallel binary arithmetic, something Atanasoff had realized years earlier. The basic design, which he first described, is now known as a von Neumann machine. It was used in the EDSAC, the first stored program computer, and is still the basis for nearly all digital computers, even now, more than half a century later. This design, and the IAS machine, built in collaboration with Herman Goldstine, has had such an enormous influence that it is worth describing briefly. Although Von Neumann’s name is always attached to this design, Goldstine and others made substantial contributions to it as well. A sketch of the architecture is given in Fig. 1-5. The von Neumann machine had five basic parts: the memory, the arithmetic logic unit, the control unit, and the input and output equipment. The memory consisted of 4096 words, a word holding 40 bits, each a 0 or a 1. Each word held either two 20-bit instructions or a 40-bit signed integer. The instructions had 8 bits devoted to telling the instruction type, and 12 bits for specifying one of the 4096 memory words. Together, the arithmetic logic unit and the control unit formed the ‘‘brain’’ of the computer. In modern computers they are combined onto a single chip called the CPU (Central Processing Unit). Inside the arithmetic logic unit was a special internal 40-bit register called the accumulator. A typical instruction added a word of memory to the accumulator

SEC. 1.2

MILESTONES IN COMPUTER ARCHITECTURE

19

Memory

Control unit

Arithmetic logic unit

Input Output

Accumulator

Figure 1-5. The original von Neumann machine.

or stored the contents of the accumulator in memory. The machine did not have floating-point arithmetic because von Neumann felt that any competent mathematician ought to be able to keep track of the decimal point (actually the binary point) in his or her head. At about the same time von Neumann was building the IAS machine, researchers at M.I.T. were also building a computer. Unlike IAS, ENIAC and other machines of its type, which had long word lengths and which were intended for heavy number crunching, the M.I.T. machine, the Whirlwind I, had a 16-bit word and was designed for real-time control. This project led to the invention of the magnetic core memory by Jay Forrester, and then eventually to the first commercial minicomputer. While all this was going on, IBM was a small company engaged in the business of producing card punches and mechanical card sorting machines. Although IBM had provided some of Aiken’s financing, it was not terribly interested in computers until it produced the 701 in 1953, long after Eckert and Mauchley’s company was number one in the commercial market with its UNIVAC computer. The 701 had 2048 36-bit words, with two instructions per word. It was the first in a series of scientific machines that came to dominate the industry within a decade. Three years later came the 704, which initially had 4096 words of core memory, 36-bit instructions, and a new innovation, floating-point hardware. In 1958, IBM began production of its last vacuum tube machine, the 709, which was basically a beefed-up 704.

1.2.3 The Second Generation—Transistors (1955–1965) The transistor was invented at Bell Labs in 1948 by John Bardeen, Walter Brattain, and William Shockley, for which they were awarded the 1956 Nobel Prize in physics. Within 10 years the transistor revolutionized computers, and by the late 1950s, vacuum tube computers were obsolete. The first transistorized

20

INTRODUCTION

CHAP. 1

computer was built at M.I.T.’s Lincoln Laboratory, a 16-bit machine along the lines of the Whirlwind I. It was called the TX-0 (Transistorized eXperimental computer 0) and was merely intended as a device to test the much fancier TX-2. The TX-2 never amounted to much, but one of the engineers working at the Laboratory, Kenneth Olsen, formed a company, Digital Equipment Corporation (DEC) in 1957 to manufacture a commercial machine much like the TX-0. It was four years before this machine, the PDP-1, appeared, primarily because the venture capitalists who funded DEC firmly believed that there was no market for computers. After all, T.J. Watson, former president of IBM, once said that the world market for computers was about four or five units. Instead, DEC mostly sold small circuit boards. When the PDP-1 finally appeared in 1961, it had 4096 words of 18-bit words and could execute 200,000 instructions/sec. This performance was half that of the IBM 7090, the transistorized successor to the 709, and fastest computer in the world at the time. The PDP-1 cost $120,000; the 7090 cost millions. DEC sold dozens of PDP-1s, and the minicomputer industry was born. One of the first PDP-1s was given to M.I.T., where it quickly attracted the attention of some of the budding young geniuses so common at M.I.T. One of the PDP-1’s many innovations was a visual display and the ability to plot points anywhere on its 512 by 512 screen. Before long, the students had programmed the PDP-1 to play spacewar, and the world had its first video game. A few years later DEC introduced the PDP-8, which was a 12-bit machine, but much cheaper than the PDP-1 ($16,000). The PDP-8 had a major innovation: a single bus, the omnibus, as shown in Fig. 1-6. A bus is a collection of parallel wires used to connect the components of a computer. This architecture was a major departure from the memory-centered IAS machine and has been adopted by nearly all small computers since. DEC eventually sold 50,000 PDP-8s, which established it as the leader in the minicomputer business. CPU

Memory

Console terminal

Paper tape I/O

Other I/O

Omnibus

Figure 1-6. The PDP-8 omnibus.

Meanwhile, IBM’s reaction to the transistor was to build a transistorized version of the 709, the 7090, as mentioned above, and later the 7094. The 7094 had a cycle time of 2 microsec and 32,536 words of 36-bit words of core memory. The 7090 and 7094 marked the end of the ENIAC-type machines, but they dominated scientific computing for years in the 1960s. At the same time IBM had become a major force in scientific computing with the 7094, it was making a huge amount of money selling a little business-oriented

SEC. 1.2

MILESTONES IN COMPUTER ARCHITECTURE

21

machine called the 1401. This machine could read and write magnetic tapes, read and punch cards, and print output almost as fast as the 7094, and at a fraction of the price. It was terrible at scientific computing, but for business record keeping it was perfect. The 1401 was unusual in that it did not have any registers, or even a fixed word length. Its memory was 4000 8-bit bytes, although later models supported up to a then-astounding 16,000 bytes. Each byte contained a 6-bit character, an administrative bit, and a bit used to indicate end-of-word. A MOVE instruction, for example, had a source and a destination address and began moving bytes from the source to the destination until it hit one with the end-of-word bit set to 1. In 1964 a tiny unknown company, Control Data Corporation (CDC), introduced the 6600, a machine that was nearly an order of magnitude faster than the mighty 7094 and every other machine in existence at the time. It was love at first sight among the number crunchers, and CDC was launched on its way to success. The secret to its speed, and the reason it was so much faster than the 7094, was that inside the CPU was a highly parallel machine. It had several functional units for doing addition, others for doing multiplication, and still another for division, and all of them could run in parallel. Although getting the most out of it required careful programming, with some work it was possible to have 10 instructions being executed at once. As if this was not enough, the 6600 had a number of little computers inside to help it, sort of like Snow White and the Seven Vertically Challenged People. This meant that the CPU could spend all its time crunching numbers, leaving all the details of job management and input/output to the smaller computers. In retrospect, the 6600 was decades ahead of its time. Many of the key ideas found in modern computers can be traced directly back to the 6600. The designer of the 6600, Seymour Cray, was a legendary figure, in the same league as Von Neumann. He devoted his entire life to building faster and faster machines, now called supercomputers, including the 6600, 7600, and Cray-1. He also invented a now-famous algorithm for buying cars: you go to the dealer closest to your house, point to the car closest to the door and say: ‘‘I’ll take that one.’’ This algorithm wastes the least time on unimportant things (like buying cars) to leave you the maximum time for doing important things (like designing supercomputers). There were many other computers in this era, but one stands out for quite a different reason and is worth mentioning: the Burroughs B5000. The designers of machines like the PDP-1, 7094, and 6600 were all totally preoccupied with the hardware, either making it cheap (DEC) or fast (IBM and CDC). Software was almost completely irrelevant. The B5000 designers took a different tack. They built a machine specifically with the intention of having it programmed in Algol 60, a forerunner of C and Java, and included many features in the hardware to ease the compiler’s task. The idea that software also counted was born. Unfortunately it was forgotten almost immediately.

22

INTRODUCTION

CHAP. 1

1.2.4 The Third Generation—Integrated Circuits (1965–1980) The invention of the silicon integrated circuit by Robert Noyce in 1958 allowed dozens of transistors to be put on a single chip. This packaging made it possible to build computers that were smaller, faster, and cheaper than their transistorized predecessors. Some of the more significant computers from this generation are described below. By 1964 IBM was the leading computer company and had a big problem with its two highly successful machines, the 7094 and the 1401: they were as incompatible as two machines could be. One was a high-speed number cruncher using parallel binary arithmetic on 36-bit registers, and the other was a glorified input/output processor using serial decimal arithmetic on variable-length words in memory. Many of its corporate customers had both and did not like the idea of having two separate programming departments with nothing in common. When the time came to replace these two series, IBM took a radical step. It introduced a single product line, the System/360, based on integrated circuits, that was designed for both scientific and commercial computing. The System/360 contained many innovations, the most important of which was that it was a family of about a half-dozen machines with the same assembly language, and increasing size and power. A company could replace its 1401 with a 360 Model 30 and its 7094 with a 360 Model 75. The Model 75 was bigger and faster (and more expensive), but software written for one of them could, in principle, run on the other. In practice, software written for a small model would run on a large model without problems, but when moving to a smaller machine, the program might not fit in memory. Still, this was a major improvement over the situation with the 7094 and 1401. The idea of machine families caught on instantly, and within a few years most computer manufacturers had a family of common machines spanning a wide range of price and performance. Some characteristics of the initial 360 family are shown in Fig. 1-7. Other models were introduced later. ############################################################################# ! ! Model 30 ! Model 40 ! Model 50 ! Model 65 ! Property #! ############################################################################ ! ! ! ! ! Relative performance 1 ! 3.5 ! 10 ! 21 ! ! #! ############################################################################ ! Cycle time (in billionths of a sec) ! 1000 ! 625 ! 500 ! 250 ! #! ############################################################################ ! ! ! ! ! ! 65,536 ! 262,144 ! 262,144 ! 524,288 ! Maximum memory (bytes) #! ############################################################################ ! ! ! ! ! ! Bytes fetched per cycle 1 ! 2 4 ! 16 ! #! ############################################################################ ! ! !!############################################################################# !! Maximum number of data channels !! 3 !! 3 4 !! 6 !! Figure 1-7. The initial offering of the IBM 360 product line.

Another major innovation in the 360 was multiprogramming, having several programs in memory at once, so that when one was waiting for input/output to complete, another could compute. This resulted in a higher CPU utilization.

SEC. 1.2

MILESTONES IN COMPUTER ARCHITECTURE

23

The 360 also was the first machine that could emulate (simulate) other computers. The smaller models could emulate the 1401, and the larger ones could emulate the 7094, so that customers could continue to run their old unmodified binary programs while converting to the 360. Some models ran 1401 programs so much faster than the 1401 itself that many customers never converted their programs. Emulation was easy on the 360 because all the initial models and most of the later models were microprogrammed. All IBM had to do was write three microprograms, for the native 360 instruction set, the 1401 instruction set, and the 7094 instruction set. This flexibility was one of the main reasons microprogramming was introduced. The 360 solved the dilemma of binary-parallel versus serial decimal with a compromise: the machine had 16 32-bit registers for binary arithmetic, but its memory was byte-oriented, like that of the 1401. It also had 1401 style serial instructions for moving variable-sized records around memory. Another major feature of the 360 was a (for that time) huge address space of 224 (16,777,216) bytes. With memory costing several dollars per byte in those days, this much memory looked very much like infinity. Unfortunately, the 360 series was later followed by the 370 series, 4300 series, 3080 series, and 3090 series, all using the same architecture. By the mid 1980s, the memory limit became a real problem, and IBM had to partially abandon compatibility when it went to 32-bit addresses needed to address the new 232 byte memory. With hindsight, it can be argued that since they had 32-bit words and registers anyway, they probably should have had 32-bit addresses as well, but at the time no one could imagine a machine with 16 million bytes of memory. Faulting IBM for this lack of vision is like faulting a modern personal computer vendor for having only 32-bit addresses. In a few years personal computers may need far more than 4 billion bytes of memory, at which time 32-bit addresses will become intolerably small. The minicomputer world also took a big step forward in the third generation with DEC’s introduction of the PDP-11 series, a 16-bit successor to the PDP-8. In many ways, the PDP-11 series was like a little brother to the 360 series just as the PDP-1 was like a little brother to the 7094. Both the 360 and PDP-11 had wordoriented registers and a byte-oriented memory and both came in a range spanning a considerable price/performance ratio. The PDP-11 was enormously successful, especially at universities, and continued DEC’s lead over the other minicomputer manufacturers.

1.2.5 The Fourth Generation—Very Large Scale Integration (1980-?) By the 1980s, VLSI (Very Large Scale Integration) had made it possible to put first tens of thousands, then hundreds of thousands, and finally millions of transistors on a single chip. This development soon led to smaller and faster

24

INTRODUCTION

CHAP. 1

computers. Before the PDP-1, computers were so big and expensive that companies and universities had to have special departments called computer centers to run them. With the advent of the minicomputer, a department could buy its own computer. By 1980, prices had dropped so low that it was feasible for a single individual to have his or her own computer. The personal computer era had begun. Personal computers were used in a very different way than large computers. They were used for word processing, spreadsheets, and numerous highly interactive applications (such as games) that the larger computers could not handle well. The first personal computers were usually sold as kits. Each kit contained a printed circuit board, a bunch of chips, typically including an Intel 8080, some cables, a power supply, and perhaps an 8-inch floppy disk. Putting the parts together to make a computer was up to the purchaser. Software was not supplied. If you wanted any, you wrote your own. Later, the CP/M operating system, written by Gary Kildall, became popular on 8080s. It was a true (floppy) disk operating system, with a file system, and user commands typed in from the keyboard to a command processor (shell). Another early personal computer was the Apple and later the Apple II, designed by Steve Jobs and Steve Wozniak in the proverbial garage. This machine was enormously popular with home users and at schools and made Apple a serious player almost overnight. After much deliberating and observing what other companies were doing, IBM, then the dominant force in the computer industry, finally decided it wanted to get into the personal computer business. Rather than design the entire machine from scratch, using only IBM parts, which would have taken far too long, IBM did something quite uncharacteristic. It gave an IBM executive, Philip Estridge, a large bag of money and told him to go somewhere far from the meddling bureaucrats at corporate headquarters in Armonk, NY, and not come back until he had a working personal computer. Estridge set up shop far from headquarters, in Boca Raton, FL, chose the Intel 8088 as his CPU, and built the IBM Personal Computer from commercial components. It was introduced in 1981 and instantly became the best-selling computer in history. IBM also did something uncharacteristic that it would later come to regret. Rather than keeping the design of the machine totally secret (or at least, guarded by a wall of patents), as it normally did, it published the complete plans, including all the circuit diagrams, in a book that it sold for $49. The idea was to make it possible for other companies to make plug-in boards for the IBM PC, to increase its flexibility and popularity. Unfortunately for IBM, since the design was now completely public and all the parts were easily available from commercial vendors, numerous other companies began making clones of the PC, often for far less money than IBM was charging. Thus an entire industry started. Although other companies made personal computers using non-Intel CPUs, including Commodore, Apple, and Atari, the momentum of the IBM PC industry

SEC. 1.2

MILESTONES IN COMPUTER ARCHITECTURE

25

was so large that the others were steamrollered. Only a few survived, and these were in niche markets. One that did survive, although barely, was the Apple Macintosh. The Macintosh was introduced in 1984 as the successor to the ill-fated Apple Lisa, which was the first computer to come with a GUI (Graphical User Interface), similar to the now-popular Windows interface. The Lisa failed because it was too expensive, but the lower-priced Macintosh introduced a year later was a huge success and inspired love and passion among its many admirers. The early personal computer market also led to the then-unheard of desire for portable computers. At that time, a portable computer made as much sense as a portable refrigerator does now. The first true portable personal computer was the Osborne-1, which at 11 kg was more of a luggable computer than a portable computer. Still, it proved that portables were possible. The Osborne-1 was a modest commercial success, but a year later Compaq brought out its first portable IBM PC clone and was quickly established as the leader in the market for portable computers. The initial version of the IBM PC came equipped with the MS-DOS operating system supplied by the then-tiny Microsoft Corporation. As Intel was able to produce increasingly powerful CPUs, IBM and Microsoft were able to develop a successor to MS-DOS called OS/2, which featured a graphical user interface, similar to that of the Apple Macintosh. Meanwhile, Microsoft also developed its own operating system, Windows, which ran on top of MS-DOS, just in case OS/2 did not catch on. To make a long story short, OS/2 did not catch on, IBM and Microsoft had a big and extremely public falling out, and Microsoft went on to make Windows a huge success. How tiny Intel and even tinier Microsoft managed to dethrone IBM, one of the biggest, richest, and most powerful corporations in the history of the world, is a parable no doubt related in great detail in business schools around the world. With the success of the 8088 in hand, Intel went on to make bigger and better versions of it. Particularly noteworthy was the 386, released in 1985, which was essentially the first Pentium. Although modern Pentiums are much faster than the 386, in terms of architecture, the modern Pentium is basically a souped-up 386. By the mid-1980s, a new development called RISC began to take over, replacing complicated (CISC) architectures with much simpler (but faster) ones. In the 1990s, superscalar CPUs began to appear. These machines could execute multiple instructions at the same time, often in a different order than they appeared in the program. We will introduce the concepts of CISC, RISC, and superscalar in Chap. 2 and discuss them at length throughout this book. Up until 1992, personal computers were either 8-bit, 16-bit, or 32-bit. Then DEC came out with the revolutionary 64-bit Alpha, a true 64-bit RISC machine that outperformed all other personal computers by a wide margin. It had a modest success, but it was almost a decade later before 64-bit machines began to catch on in a big way, and then mostly as high-end servers.

26

INTRODUCTION

CHAP. 1

1.2.6 The Fifth Generation—Invisible Computers In 1981, the Japanese government announced that they were planning to spend $500 million to help Japanese companies develop fifth-generation computers, which would be based on artificial intelligence and represent a quantum leap over ‘‘dumb’’ fourth-generation computers. Having seen Japanese companies take over the market in many industries, from cameras to stereos to televisions, American and European computer makers went from 0 to full panic in a millisecond, demanding government subsidies and more. Despite lots of fanfare, the Japanese fifth-generation project basically failed and was quietly abandoned. In a sense, it was like Babbage’s analytical engine—a visionary idea but so far ahead of its time that the technology for actually building it was nowhere in sight. Nevertheless, what might be called the fifth generation did happen, but in an unexpected way: computers shrunk. The Apple Newton, released in 1993, showed that a computer could be built in a package no bigger than a portable audio cassette player. The Newton used handwriting for user input, which proved to be a big stumbling block, but later machines of this class, now called PDAs (Personal Digital Assistants), had improved user interfaces and became very popular. Many of these now have almost as much computing power as personal computers from a few years earlier. But even the PDAs are not really revolutionary. Even more important are the ‘‘invisible’’ computers, which are embedded into appliances, watches, bank cards, and numerous other devices (Bechini et al., 2004). These processors allow increased functionality and lower cost in a wide variety of applications. Whether these chips form a true generation is debatable (they have been around since the 1970s), but they are revolutionizing how thousands of appliances and other devices work. They are already starting to have a major impact on the world and their influence will increase rapidly in the coming years. One unusual aspects of these embedded computers is that the hardware and software are often codesigned (Henkel et al., 2003). We will come back to them later in this book. If we see the first generation as vacuum tube machines (e.g. ENIAC), the second generation as transistor machines (e.g., the IBM 7094), the third generation as early integrated circuit machines (e.g., the IBM 360), and the fourth generation as personal computers (e.g., the Intel CPUs), the real fifth generation is more a paradigm shift than a specific new architecture. In the future, computers will be everywhere and embedded in everything—indeed, invisible. They will be part of the framework of daily life, opening doors, turning on lights, dispensing money, and thousands of other things. This model, devised by the late Mark Weiser was originally called ubiquitous computing but the term pervasive computing is also used frequently now (Weiser, 2002). It will change the world as profoundly as the industrial revolution did. We will not discuss it further in this book, but for more information about it, see (Lyytinen and Yoo, 2002; Saha and Mukherjee, 2003; and Sakamura, 2002).

SEC. 1.3

27

THE COMPUTER ZOO

1.3 THE COMPUTER ZOO In the previous section, we gave a very brief history of computer systems. In this one we will look at the present and gaze toward the future. Although personal computers are the best known computers, there are other kinds of machines around these days, so it is worth taking a brief look at what else is out there.

1.3.1 Technological and Economic Forces The computer industry is moving ahead like no other. The primary driving force is the ability of chip manufacturers to pack more and more transistors per chip every year. More transistors, which are tiny electronic switches, means larger memories and more powerful processors. Gordon Moore, co-founder and former chairman of Intel, once joked that if aviation technology had moved ahead as fast as computer technology, an airplane would cost $500 and circle the earth in 20 minutes on 5 gallons of fuel. However, it would be the size of a shoebox. Specifically, while preparing a speech for an industry group, Moore noticed that each new generation of memory chips was being introduced 3 years after the previous one. Since each new generation had four times as much memory as its predecessor, he realized that the number of transistors on a chip was increasing at a constant rate and predicted this growth would continue for decades to come. This observation has become known as Moore’s law. Today, Moore’s law is often expressed as the number of transistors doubling every 18 months. Note that this is equivalent to about a 60 percent increase in transistor count per year. The sizes of the memory chips and their dates of introduction shown in Fig. 1-8 confirm that Moore’s law has held for over three decades. 512M

1,000,000,000 100,000,000

16M

10,000,000

1M

Transistors

1,000,000 100,000

4K

10,000 1,000

1K

64K

64M

256M

4M

256K

16K

100 10 1 1965

1970

1975

1980

1985

1990

1995

2000

Figure 1-8. Moore’s law predicts a 60-percent annual increase in the number of transistors that can be put on a chip. The data points given in this figure are memory sizes, in bits.

28

INTRODUCTION

CHAP. 1

Of course, Moore’s law is not a law at all, but simply an empirical observation about how fast solid state physicists and process engineers are advancing the state-of-the-art, and a prediction that they will continue at the same rate in the future. Some industry observers expect Moore’s law to continue for at least another decade, maybe longer. At that point transistors will consist of too few atoms to be reliable, although advances in quantum computing may conceivable change that (Oskin et al., 2002). However, other observers expect energy dissipation, current leakage, and other effects to kick in earlier and cause serious problems that need to be solved (Bose, 2004; Kim et al., 2003). Moore’s law has created what economist’s call a virtuous circle. Advances in technology (transistors/chip) lead to better products and lower prices. Lower prices lead to new applications (nobody was making video games for computers when computers cost $10 million each). New applications lead to new markets and new companies springing up to take advantage of them. The existence of all these companies leads to competition, which in turn, creates economic demand for better technologies with which to beat the others. The circle is then round. Another factor driving technological improvement is Nathan’s first law of software (due to Nathan Myhrvold, a former top Microsoft executive). It states: ‘‘Software is a gas. It expands to fill the container holding it.’’ Back in the 1980s, word processing was done with programs like troff (still used for this book). Troff occupies kilobytes of memory. Modern word processors occupy megabytes of memory. Future ones will no doubt require gigabytes of memory. (To a first approximation, the prefixes kilo, mega, and giga mean thousand, million, and billion, respectively, but see Sec. 1.5 for details.) Software that continues to acquire features (not unlike boats that continue to acquire barnacles) creates a constant demand for faster processors, bigger memories, and more I/O capacity. While the gains in transistors per chip have been dramatic over the years, the gains in other computer technologies have been hardly less so. For example, the IBM PC/XT was introduced in 1982 with a 10-megabyte hard disk. Twenty years later, 100-gigabyte hard disks were common on the PC/XT’s successors. This improvement of four orders of magnitude in 20 years represents an annual capacity increase of 58 percent. However, measuring disk improvement is trickier, since there are other parameters besides capacity, such as data rate, seek time, and price. Nevertheless, almost any metric will show that the price/performance has increased since 1982 by at least 50 percent per year. These enormous gains in disk performance, coupled with the fact that the dollar volume of disks shipped from Silicon Valley has exceeded that of CPU chips led Al Hoagland to suggest that the place was named wrong: it should have been called Iron Oxide Valley (since this is the recording medium used on disks). Another area that has seen spectacular gains has been telecommunication and networking. In less than two decades, we have gone from 300 bit/sec modems, to analog modems at 56,000 bits/sec to fiber-optic networks at 1012 bits/sec. Fiberoptic transatlantic telephone cables, such as TAT-12/13, cost about $700 million,

SEC. 1.3

THE COMPUTER ZOO

29

last for 10 years, and can carry 300,000 simultaneous calls, which comes to under one cent for a 10-minute intercontinental call. Optical communication systems running at 1012 bits/sec) over distances exceeding 100 km without amplifiers have been proven feasible. The exponential growth of the Internet hardly needs comment here.

1.3.2 The Computer Spectrum Richard Hamming, a former researcher at Bell Labs, once observed that a change of an order of magnitude in quantity causes a change in quality. Thus a racing car that can go 1000 km/hour in the Nevada desert is a fundamentally different kind of machine than a normal car that goes 100 km/hour on a highway. Similarly, a 100-story skyscraper is not just a scaled up 10-story apartment building. And with computers, we are not talking about factors of 10, but over the course of three decades, factors of a million. The gains afforded by Moore’s law can be used in several different ways. One way is to build increasingly powerful computers at constant price. Another approach is to build the same computer for less and less money every year. The computer industry has done both of these and more, resulting in a wide variety of computers available now. A very rough categorization of current computers is given in Fig. 1-9. ######################################################################## ! ! Price ($) ! ! Type Example application #! ####################################################################### ! ! ! Disposable computer 0.5 ! Greeting cards ! ! #! ####################################################################### ! Microcontroller ! ! Watches, cars, appliances ! 5 #! ####################################################################### ! ! ! ! ! Home video games ! Game computer 50 #! ####################################################################### ! ! ! ! Personal computer 500 Desktop or notebook computer !######################################################################## ! ! ! !######################################################################## ! ! Server 5K ! Network server ! ! ! ! Collection of Workstations 50–500K Departmental minisupercomputer !######################################################################## ! ! ! ! Mainframe ! ! Batch data processing in a bank ! 5M !######################################################################## ! ! ! Figure 1-9. The current spectrum of computers available. The prices should be taken with a grain (or better yet, a metric ton) of salt.

In the following sections we will examine each of these categories and discuss their properties briefly.

1.3.3 Disposable Computers At the bottom end, we find single chips glued to the inside of greeting cards for playing ‘‘Happy Birthday,’’ ‘‘Here Comes the Bride,’’ or some equally appalling ditty. The author has not yet spotted a condolence card that plays a funeral dirge, but having now released this idea into the public domain, he expects it

30

INTRODUCTION

CHAP. 1

shortly. To anyone who grew up with multimillion-dollar mainframes, the idea of disposable computers makes about as much sense as disposable aircraft. However, disposable computers are here to stay. Probably the most important development in the area of throwaway computers is the RFID (Radio Frequency IDentification) chip. It is now possible to manufacture, for a few cents, batteryless RFID chips smaller than 0.5 mm on edge that contain a tiny radio transponder and a built-in unique 128-bit number. When pulsed from an external antenna, they are powered by the incoming radio signal long enough to transmit their number back to the antenna. While the chips are tiny, their implications are certainly not. Let us start with a mundane application: removing bar codes from products. Experimental trials have already been held in which products in stores have RFID chips (instead of bar codes) attached by the manufacturer. The customer selects her products, puts them in a shopping cart, and just wheels them out of the store, bypassing the checkout counter. At the store’s exit, a reader with an antenna sends out a signal asking each product to identify itself, which it does by a short wireless transmission. The customer is also identified by a chip on her bank card or credit card. At the end of the month, the store sends the customer an itemized bill for this month’s purchases. If the customer does not have a valid RFID bank or credit card, an alarm is sounded. Not only does this system eliminate the need for cashiers and the corresponding wait in line, but it also serves as an antitheft system because hiding a product in a pocket or bag has no effect. An interesting property of this system is that while bar codes identify the product type, they do not identify the specific item. With 128 bits available, RFID chips do. As a consequence, every package of, say, aspirins, on a supermarket shelf will have a different RFID code. This means that if a drug manufacturer discovers a manufacturing defect in a batch of aspirins after they have been shipped, supermarkets all over the world can be told to sound the alarm when a customer buys any package whose RFID number lies in the affected range, even if the purchase happens in a distant country months later. Aspirins not in the defective batch will not sound the alarm. But labeling packages of aspirins, cookies, and dog biscuits is only the start. Why stop at labeling the dog biscuits when you can label the dog? Pet owners are already asking veterinarians to implant RFID chips in their animals, to allow them to be traced if they are stolen or lost. Farmers want their livestock tagged as well. The obvious next step is for nervous parents to ask their pediatrician to implant RFID chips in their children in case they get stolen or lost. While we are at it, why not have hospitals put them in all newborns to avoid mixups at the hospital. Governments and the police can no doubt think of many good reasons for tracking all citizens all the time. By now, the ‘‘implications’’ of RFID chips alluded to above may be getting a bit clearer. Another (slightly less controversial) application of RFID chips is vehicle tracking. When a string of railroad cars with embedded RFID chips passes by a

SEC. 1.3

THE COMPUTER ZOO

31

reader, the computer attached to the reader then has a list of which cars passed by. This system makes it easy to keep track of the location of all railroad cars, which helps suppliers, their customers, and the railroads. A similar scheme can be applied to trucks. For cars, the idea can be used for collecting tolls electronically. Airline baggage systems and many other package transport systems can also use RFID chips. An experimental system tested at Heathrow airport in London allowed arriving passengers to remove the lugging from their luggage. Bags carried by passengers purchasing this service were tagged with RFID chips, routed separately within the airport, and delivered directly to the passengers’ hotels. Other uses of RFID chips include having cars arriving at the painting station of the assembly line specify what color they are supposed to be, studying animal migrations, having clothes tell the washing machine what temperature to use, and many more. Some chips may be integrated with sensors so that the low-order bits may contain the current temperature, pressure, humidity or other environmental variable. Advanced RFID chips also contain permanent storage. This capability led the European Central Bank to make a decision to put RFID chips in euro banknotes in the coming years. The chips would record where they have been. Not only would this make counterfeiting euro notes virtually impossible, but it would make tracing kidnapping ransoms, the loot taken from robberies, and laundered money much easier to track and possibly remotely invalidated. When cash is no longer anonymous, standard police procedure in the future may be to check out where the suspect’s money has been recently. Who needs to implant chips in people when their wallets are full of them? Again, when the public learns about what RFID chips can do, there is likely to be some public discussion about the matter. The technology used in RFID chips is developing rapidly. The smallest ones are passive (not internally powered) and only capable of transmitting their unique numbers when queried. However, larger ones are active, can contain a small battery and a primitive computer, and are capable of doing some calculations. Smart cards used in financial transactions fall into this category. RFID chips differ not only in being active or passive, but also in the range of radio frequencies they respond to. Those operating at low frequences have a limited data rate but can be sensed at great distances from the antenna. Those operating at high frequencies have a higher data rate and a shorter range. The chips also differ in other ways and are being improved all the time. The Internet is full of information about RFID chips, with www.rfid.org being one good starting point.

1.3.4 Microcontrollers Next up the ladder we have computers that are embedded inside devices that are not sold as computers. The embedded computers, sometimes called microcontrollers, manage the devices and handle the user interface. Microcontrollers

32

INTRODUCTION

CHAP. 1

are found in a large variety of different devices, including the following. Some examples of each category are given in parentheses. 1. Appliances (clock radio, washer, dryer, microwave, burglar alarm). 2. Communications gear (cordless phone, cell phone, fax, pager). 3. Computer peripherals (printer, scanner, modem, CD ROM drive). 4. Entertainment devices(VCR, DVD, stereo, MP3 player, set top box). 5. Imaging devices (TV, digital camera, camcorder, lens, photocopier). 6. Medical devices (X-ray, MRI, heart monitor, digital thermometer). 7. Military weapon systems (cruise missile, ICBM, torpedo). 8. Shopping devices (vending machine, ATM, cash register). 9. Toys (talking doll, game console, radio-controlled car or boat). A high-end car could easily contain 50 microcontrollers, running subsystems including the antilock brakes, fuel injection, radio, and GPS. A jet plane could easily have 200 or more of them. A family might easily own several hundred computers without knowing it. Within a few years, practically everything that runs on electricity or batteries will contain a microcontroller. The numbers of microcontrollers sold every year dwarfs all other kinds of computers except disposable computers by orders of magnitude. While RFID chips are minimal systems, microcontrollers are small, but complete, computers. Each microcontroller has a processor, memory, and I/O capability. The I/O capability usually includes sensing the device’s buttons and switches and controlling the device’s lights, display, sound. motors In most cases, the software is hardwired into the chip in the form of a read-only memory created when the microcontroller is manufactured. Microcontrollers come in two general types: general purpose and special purpose. The former are just small, but ordinary computers; the latter have an architecture and instruction set tuned to some specific application, for example, multimedia. Microcontrollers come in 4-bit, 8bit, 16-bit, and 32-bit versions. However, even the general-purpose microcontrollers differ from standard PCs in important ways. First, they are extremely cost sensitive. A company buying millions of units may make the choice based on a 1 cent price difference per unit. This constraint makes microcontroller manufacturers make architectural choices based on manufacturing costs much more than on chips costing hundreds of dollars. Although microcontroller prices vary greatly depending on how many bits wide they are, how much and what kind of memory they have, and other factors; to get an idea, an 8-bit microcontroller purchased in large enough volume can probably be had for as little as 10 cents per unit. This price is what makes it possible to put a computer inside a $9.95 clock radio.

SEC. 1.3

THE COMPUTER ZOO

33

Second, virtually all microcontrollers operate in real time. They get a stimulus and are expected to give an instantaneous response. For example, when the user presses a button, often a light goes on, and there should not be any delay between the button being pressed and the light going on. The need to operate in real time often has impact on the architecture. Third, embedded systems often have physical constraints in terms of size, weight, battery consumption, and other electrical and mechanical limits. The microcontrollers used in them have to be designed with these restrictions in mind.

1.3.5 Game Computers A step up are the video game machines. They are normal computers, with special graphics and sound capability, but limited software and little extensibility. They started out as low-end CPUs for playing simple action games like ping pong on TV sets. Over the years they have evolved into far more powerful systems, rivaling or even outperforming personal computers in certain dimensions. To get an idea of what is inside a game computer, consider the specifications of three popular products. First, the Sony PlayStation 2. It contains a 295-MHz 128-bit proprietary CPU (called the Emotion Engine), which is based on the MIPS IV RISC CPU. The PlayStation 2 also contains 32 MB of RAM, a 160-MHz custom graphics chip, a 48-channel custom audio chip, and a DVD player. Second, the Microsoft XBOX. It contains a 733-MHz Intel Pentium III with 64 MB of RAM, a 300-MHz custom graphics chip, a 256-channel custom audio chip, a DVD player and an 8-GB hard disk. Third, the Nintendo GameCube. It contains a 485-MHz 32-bit custom CPU (called the Gekko) derived from the IBM PowerPC RISC CPU, 24 MB of RAM, a 200-MHz custom graphics chip, a 64channel audio chip, and a proprietary 1.5 gigabyte optical disk. While these machines are not quite as powerful as personal computers produced in the same time period, they are not that far behind, and in some ways are ahead (e.g., the 128-bit CPU in the PlayStation 2 is wider than the CPU in any PC, although the clock speed is much lower). The main difference between a game machine and a PC is not so much the CPU as it is the fact that game machines are closed systems. Users may not expand them with plug-in cards, although USB or FireWire interfaces are sometimes provided. Also, and perhaps most important, game machines are carefully optimized for a single application area: highly interactive 3D games with high quality stereo audio. Everything else is secondary. These hardware and software restrictions, slow clock speeds, small memories, absence of a high-resolution monitor, and (usually) absence of a hard disk make it possible to build and sell these machines more cheaply than personal computers. Despite these restrictions, millions of game machines have been sold. The same companies that make the main game machines also make portable game machines that are handheld and run on batteries. These are closer to the embedded systems we discussed above than to personal computers though.

34

INTRODUCTION

CHAP. 1

1.3.6 Personal Computers Next, we come to the personal computers that most people think of when they hear the term ‘‘computer.’’ These include desktop and notebook models. They usually come with hundreds of megabytes of memory, a hard disk holding around 100 gigabytes of data, a CD-ROM/DVD drive, modem, sound card, network interface, high-resolution monitor, and other peripherals. They have elaborate operating systems, many expansion options, and a huge range of available software. Some people reserve the term ‘‘PC’’ for those machines that have an Intel CPU and use ‘‘workstation’’ for those powered by a high-end RISC chip, such as the Sun UltraSPARC. Conceptually, however, there is little difference between them. The heart of every personal computer is a printed circuit board at the bottom of the case. It usually contains the CPU. memory, various I/O devices (such as a sound chip and possibly a modem), as well as interfaces to the keyboard, mouse, disk, network, etc., and some expansion slots. A picture of one of these circuit boards is given in Fig. 1-10. Notebook computers are basically PCs in a smaller package. The use the same hardware components, but manufactured in smaller sizes. They also run the same software as desktop PCs. Yet another closely related machine type is the PDA. While these are even smaller than notebook computers, each one has a CPU, memory, keyboard, display, and most of the other features of a personal computer in miniature. Since most readers are probably quite familiar with personal computers, additional introductory material is hardly needed.

1.3.7 Servers Beefed-up personal computers or workstations are often used as network servers, both for local area networks (typically within a single company), and for the Internet. These come in single-processor and multiple-processor configurations, have gigabytes of memory, hundreds of gigabytes of hard disk space, and high-speed networking capability. Some of them can handle thousands of transactions per second. Architecturally, however, a single-processor server is not really very different from a single-processor personal computer. It is just faster, bigger, has more disk space and possibly a faster network connection. Servers run the same operating systems as personal computers, typically some flavor of UNIX or Windows.

1.3.8 Collections of Workstations Due to almost continuous improvements in the price/performance ratio of workstations and personal computers, in recent years system designers have begun connecting large numbers of them together to form COWs (Clusters of

SEC. 1.3

THE COMPUTER ZOO

35

Figure 1-10. A printed circuit board is at the heart of every personal computer. This figure is a photograph of the Intel D875PBZ board. The photograph is copyrighted by the Intel Corporation, 2003 and used by permission. 1. Pentium 4 socket 2. 875P Support chip 3. Memory sockets 4. AGP connector

5. Disk interface 6. Gigabit Ethernet 7. Five PCI slots

8. USB 2.0 ports 9. Cooling technology 10. BIOS

Workstations), or sometimes just clusters. They consist of standard personal computers or workstations connected by gigabit/sec networks, and running special software that allow all the machines to work together on a single problem, often in science or engineering. Normally they are what are called COTS (Commodity Off The Shelf) computers that anyone can just buy from a normal PC vendor. The main addition is high-speed networking, but sometimes that is also a standard commercial network card too. Clusters scale easily, from a handful of machines

36

INTRODUCTION

CHAP. 1

to thousands of them. Usually, the amount of money available is the limiting factor. Due to their low component price, individual departments can now own such machines. Another use for a COW is as an Internet Web server. When a Website expects thousands of requests per second for its pages, the most economical solution is often a cluster with hundreds, or even thousands, of servers. The incoming requests are then sprayed among the servers to allow them to be processed in parallel. When used this way, a COW is often called a server farm.

1.3.9 Mainframes Now we come to the mainframes: room-sized computers that hark back to the 1960s. In many cases, these machines are the direct descendants of IBM 360 mainframes acquired decades ago. For the most part, they are not much faster than powerful servers, but they always have more I/O capacity and are often equipped with vast disk farms, often holding thousands of gigabytes of data. While expensive, they are often kept running due to the immense investment in software, data, operating procedures, and personnel that they represent. Many companies find it cheaper to just pay a few million dollars once in a while for a new one, than to even contemplate the effort required to reprogram all their applications for smaller machines. It is this class of computer that led to the now-infamous Year 2000 problem, which was caused by COBOL programmers in the 1960s and 1970s representing the year as two decimal digits (in order to save memory). They never envisioned their software lasting three or four decades. While the predicted disaster never occurred due to a huge amount of work put into fixing the problem, many companies have repeated the same mistake by simply adding two more digits to the year. The author hereby predicts the end of civilization as we know it at midnight on Dec. 31, 9999, when 8000 years worth of old COBOL programs crash simultaneously. In addition to their use for running 30-year-old legacy software, the Internet has breathed new life into mainframes in recent years. They have found a new niche as powerful Internet servers, for example, by handling massive numbers of e-commerce transactions per second, particularly in businesses where huge data bases are required. Although the focus of this book is on PCs, servers, and microcontrollers, we will look at mainframes a bit more in Chap. 5. Up until recently, there was another category of computers even more powerful than mainframes: supercomputers. They had enormously fast CPUs, many gigabytes of main memory, and very fast disks and networks. They were used for massive scientific and engineering calculations such as simulating colliding galaxies, synthesizing new medicines, or modeling the flow of air around an airplane wing. However, in recent years, COWs have come to offer as much computing power at much lower prices, and the true supercomputers are now a dying breed.

SEC. 1.4

EXAMPLE COMPUTER FAMILIES

37

1.4 EXAMPLE COMPUTER FAMILIES In this book we will focus on three kinds of computers: personal computers. servers, and embedded computers. Personal computers are of interest because every reader has undoubtedly used one. Servers are of interest because they run all the services on the Internet. Finally, embedded computers are invisible to their users but control cars, televisions, microwave ovens, washing machines, and practically every other electrical device costing more than $50. In this section we will give a brief introduction to the three computers that will be used as examples in the rest of the book, one in each of these three categories. They are the Pentium 4, the UltraSPARC III, and the 8051.

1.4.1 Introduction to the Pentium 4 In 1968, Robert Noyce, inventor of the silicon integrated circuit, Gordon Moore, of Moore’s law fame, and Arthur Rock, a San Francisco venture capitalist, formed the Intel Corporation to make memory chips. In its first year of operation, Intel sold only $3000 worth of chips, but business has picked up since then. In the late 1960s, calculators were large electromechanical machines the size of a modern laser printer and weighing 20 kg. In Sept. 1969, a Japanese company, Busicom, approached Intel with a request for it to manufacture 12 custom chips for a proposed electronic calculator. The Intel engineer assigned to this project, Ted Hoff, looked at the plan and realized that he could put a 4-bit generalpurpose CPU on a single chip that would do the same thing and be simpler and cheaper as well. Thus in 1970, the first single-chip CPU, the 2300-transistor 4004 was born (Faggin et al., 1996). It is worth noting that neither Intel nor Busicom had any idea what they had just done. When Intel decided that it might be worth a try to use the 4004 in other projects, it offered to buy back all the rights to the new chip from Busicom by returning the $60,000 Busicom had paid Intel to develop it. Intel’s offer was quickly accepted, at which point it began working on an 8-bit version of the chip, the 8008, introduced in 1972. The Intel family, starting with the 4004 and 8008 is shown in Fig. 1-11. Intel did not expect much demand for the 8008, so it set up a low-volume production line. Much to everyone’s amazement, there was an enormous amount of interest, so Intel set about designing a new CPU chip that got around the 8008’s limit of 16 kilobytes of memory (imposed by the number of pins on the chip). This design resulted in the 8080, a small, general-purpose CPU, introduced in 1974. Much like the PDP-8, this product took the industry by storm and instantly became a mass market item. Only instead of selling thousands, as DEC had, Intel sold millions. In 1978 came the 8086, a genuine 16-bit CPU on a single chip. The 8086 was designed to be similar to the 8080, but it was not completely compatible with the

38

INTRODUCTION

CHAP. 1

######################################################################################### ! Date ! ! ! Memory ! ! Chip MHz Transistors Notes !######################################################################################### ! ! ! ! ! ! 4004 0.108 ! 2300 ! 640 ! First microprocessor on a chip !######################################################################################### ! 4/1971 ! ! ! 8008 ! 4/1972 ! ! ! 16 KB ! First 8-bit microprocessor ! 0.108 3500 !######################################################################################### ! ! ! ! ! ! !######################################################################################### ! 4/1974 ! 8080 2 ! 6000 ! 64 KB ! First general-purpose CPU on a chip ! ! ! ! ! ! ! ! 8086 5–10 ! 29,000 ! 1 MB ! First 16-bit CPU on a chip !######################################################################################### ! 6/1978 ! ! !######################################################################################### ! 6/1979 ! ! 8088 5–8 ! 29,000 ! 1 MB ! Used in IBM PC ! ! ! ! ! ! ! 80286 8–12 ! 134,000 ! 16 MB ! Memory protection present !######################################################################################### ! 2/1982 ! ! ! 80386 ! 10/1985 ! ! 275,000 ! ! First 32-bit CPU ! 16–33 4 GB !######################################################################################### ! ! ! ! ! ! 80486 25–100 ! 1.2M ! 4 GB ! Built-in 8-KB cache memory !######################################################################################### ! 4/1989 ! ! ! Pentium ! 3/1993 ! ! ! ! Two pipelines; later models had MMX ! 60–233 3.1M 4 GB !######################################################################################### ! ! ! ! ! ! !######################################################################################### ! Pentium Pro ! 3/1995 ! 150–200 ! 5.5M ! 4 GB ! Two levels of cache built in ! ! ! ! ! ! ! Pentium II ! 5/1997 ! 233–450 ! 7.5M ! 4 GB ! Pentium Pro plus MMX instructions ! !######################################################################################### ! ! Pentium III ! 2/1999 ! 650–1400 ! 4 GB ! SSE Instructions for 3D graphics 9.5M ! # ######################################################################################## ! ! ! ! ! ! ! Pentium 4 ! 11/2000 !1300–3800 ! 42M ! 4 GB ! Hyperthreading; more SSE instructions ! !######################################################################################### !

Figure 1-11. The Intel CPU family. Clock speeds are measured in MHz (megahertz) where 1 MHz is 1 million cycles/sec.

8080. The 8086 was followed by the 8088, which had the same architecture as the 8086, and ran the same programs but had an 8-bit bus instead of a 16-bit bus, making it both slower and cheaper than the 8086. When IBM chose the 8088 as the CPU for the original IBM PC, this chip quickly became the personal computer industry standard. Neither the 8088 nor the 8086 could address more than 1 megabyte of memory. By the early 1980s this became more and more of a serious problem, so Intel designed the 80286, an upward compatible version of the 8086. The basic instruction set was essentially the same as that of the 8086 and 8088, but the memory organization was quite different, and rather awkward, due to the requirement of compatibility with the older chips. The 80286 was used in the IBM PC/AT and in the midrange PS/2 models. Like the 8088, it was a huge success, mostly because people viewed it as a faster 8088. The next logical step was a true 32-bit CPU on a chip, the 80386, brought out in 1985. Like the 80286, this one was more-or-less compatible with everything back to the 8080. Being backward compatible was a boon to people for whom running old software was important, but a nuisance to people who would have preferred a simple, clean, modern architecture unencumbered by the mistakes and technology of the past. Four years later the 80486 came out. It was essentially a faster version of the 80386 that also had a floating-point unit and 8 kilobytes of cache memory on chip. Cache memory is used to hold the most commonly used memory words inside or close to the CPU, to avoid (slow) accesses to main memory. The 80486 also had

SEC. 1.4

EXAMPLE COMPUTER FAMILIES

39

built-in multiprocessor support, to allow manufacturers to build systems containing multiple CPUs sharing a common memory. At this point, Intel found out the hard way (by losing a trademark infringement lawsuit) that numbers (like 80486) cannot be trademarked, so the next generation got a name: Pentium (from the Greek word for five, πεντε). Unlike the 80486, which had one internal pipeline, the Pentium had two of them, which helped make it twice as fast (we will discuss pipelines in detail in Chap. 2). Later in the production run, Intel added special MMX (MultiMedia eXtension) instructions. These instructions were intended to speed up computations required to process audio and video, making the addition of special multimedia coprocessors unnecessary. When the next generation appeared, people who were hoping for the Sexium (sex is Latin for six) were disappointed. The name Pentium was now so well known that the marketing people wanted to keep it, and the new chip was called the Pentium Pro. Despite the small name change from its predecessor, this processor represented a major break with the past. Instead of having two or more pipelines, the Pentium Pro had a very different internal organization and could execute up to five instructions at a time. Another innovation found in the Pentium Pro was a two-level cache memory. The processor chip itself had 8 kilobytes of fast memory to hold commonly-used instructions and another 8 kilobytes of fast memory to hold commonly-used data. In the same cavity within the Pentium Pro package (but not on the chip itself) was a second cache memory of 256 kilobytes. Although the Pentium Pro had a big cache, it lacked the MMX instructions (because Intel was unable to manufacture such a large chip with acceptable yields). When the technology improved enough to get both the MMX instructions and the cache on one chip, the combined product was released as the Pentium II. Next, yet more multimedia instructions, called SSE (Streaming SIMD Extensions), were added for enhanced 3D graphics (Raman et al., 2000). The new chip was dubbed the Pentium III, but internally it was essentially a Pentium II. The next Pentium was based on a different internal architecture. To celebrate this event, Intel switched from Roman numerals to Arabic numbers and called it the Pentium 4. As usual, the Pentium 4 was faster than all its predecessors. The 3.06 GHz version also introduced an intriguing new feature—hyperthreading. This feature allowed programs to split their work into two threads of control which the Pentium 4 could run in parallel, speeding up execution. In addition, another batch of SSE instructions was added to speed up audio and video processing even more. A photograph of the Pentium 4 chip is given in Fig. 1-12. In reality, it is about 16.0 mm × 13.5 mm, an extremely large chip. In addition to the mainline desktop CPUs discussed above, Intel has manufactured variants of some of the Pentium chips for special markets. In early 1998, Intel introduced a new product line called the Celeron, which was basically a low-price, low-performance version of the Pentium 2 intended for low-end PCs.

40

INTRODUCTION

CHAP. 1

Figure 1-12. The Pentium 4 chip. The photograph is copyrighted by the Intel Corporation, 2003 and used by permission.

Since the Celeron has the same architecture as the Pentium 2, we will not discuss it further in this book. In June 1998, Intel introduced a special version of the Pentium 2 for the upper end of the market. This processor, called the Xeon, had a larger cache, a faster bus, and better multiprocessor support, but was otherwise a normal Pentium 2, so we will not discuss it separately either. The Pentium III also had a Xeon version. In Nov. 2000, Intel released the Pentium 4, which ran the same programs as the Pentium III and Xeon, but internally was a completely new design. The 3.06 GHz version of the Pentium 4 introduced hyperthreading, a subject we will discuss in Chap. 8. 1n 2003, Intel introduced the Pentium M (as in Mobile), a chip designed for notebook computers. This chip was part of the Centrino architecture, whose goals were lower power consumption for longer battery lifetime; smaller, lighter, computers; and built-in wireless networking capability using the IEEE 802.11 (WiFi) standard. Intel intends to introduce chip sets for other specific applications in the future, for example, home entertainment devices and IEEE 802.16 (WiMax) notebooks. All the Intel chips are backward compatible with their predecessors back as far as the 8086. In other words, a Pentium 4 can run old 8086 programs without

SEC. 1.4

41

EXAMPLE COMPUTER FAMILIES

modification. This compatibility has always been a design requirement for Intel, to allow users to maintain their existing investment in software. Of course, the Pentium 4 is three orders of magnitude more complex than the 8086, so it can do quite a few things that the 8086 could not do. These piecemeal extensions have resulted in an architecture that is not as elegant as it might have been had someone given the Pentium 4 architects 42 million transistors and instructions to start all over again. It is interesting to note that although Moore’s law has long been associated with the number of bits in a memory, it applies equally well to CPU chips. By plotting the transistor counts given in Fig. 1-12 against the date of introduction of each chip on a semilog scale, we see that Moore’s law holds here too. This graph is given in Fig. 1-13. 100M 10M

Moore's law

Transistors

1M

80286

100K 10K

4004 1K 8008

80486 80386

8086 8080

Pentium

Pentium 4 Pentium III Pentium II

Pentium Pro

8088

100 10 1 1970 1972 1974 1976 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 Year of introduction

Figure 1-13. Moore’s law for (Intel) CPU chips.

While Moore’s law will probably continue to hold for some years to come, another problem is starting to overshadow it: heat dissipation. Smaller transistors make it possible to run at higher clock frequencies, which requires higher using a higher voltage. Power consumed and heat dissipated is proportional to the square of the voltage, so going faster means having more heat to get rid of. At 3.6 GHz, the Pentium 4 consumes 115 watts of power. That means it gets about as hot as a 100-watt light bulb. Speeding up the clock makes the problem worse. In November 2004, Intel canceled the 4-GHz Pentium 4 due to problems dissipating the heat. Large fans can help but the noise they make is not popular with users and water cooling, while used on large mainframes, is not an option for desktop machines (and even less so for notebook computers). As a consequence, the once-relentless march of the clock may be temporarily stymied, at least until

42

INTRODUCTION

CHAP. 1

Intel’s engineers figure out how to get rid of all the heat generated in an efficient way. Instead, Intel’s future plans call for putting two CPUs on a single chip, along with large shared cache. Because of the way power consumption is related to voltage and clock speed, two CPUs on a chip consumes far less power than one CPU at the twice the speed. As a consequence, the gain offered by Moore’s law may be increasingly exploited in the future to include larger and larger on-chip caches, rather than higher and higher clock speeds (because memory does not consume much power).

1.4.2 Introduction to the UltraSPARC III In the 1970s, UNIX was popular at universities, but no personal computers ran UNIX, so UNIX -lovers had to use (often overloaded) timeshared minicomputers

such as the PDP-11 and VAX. In 1981, a German Stanford graduate student, Andy Bechtolsheim, who was frustrated at having to go to the computer center to use UNIX, decided to solve this problem by building himself a personal UNIX workstation out of off-the-shelf parts. He called it the SUN-1 (Stanford University Network). Bechtolsheim soon attracted the attention of Vinod Khosla, a 27-year-old Indian who had a burning desire to retire as a millionaire by age 30. Khosla convinced Bechtolsheim to form a company to build and sell Sun workstations. Khosla then hired Scott McNealy, another Stanford graduate student, to head manufacturing. To write the software, they hired Bill Joy, the principle architect of Berkeley UNIX. The four of them founded Sun Microsystems in 1982. Sun’s first product, the Sun-1, which was powered by a Motorola 68020 CPU, was an instant success, as were the follow-up Sun-2 and Sun-3 machines, which also used Motorola CPUs. Unlike other personal computers of the day, these machines were far more powerful (hence the designation ‘‘workstation’’) and were designed from the start to be run on a network. Each Sun workstation came equipped with an Ethernet connection and with TCP/IP software for connecting to the ARPANET, the forerunner of the Internet. By 1987, Sun, now selling half a billion dollars a year worth of systems, decided to design its own CPU, basing it upon a revolutionary new design from the University of California at Berkeley (the RISC II). This CPU, called the SPARC (Scalable Processor ARChitecture), formed the basis of the Sun-4 workstation. Within a short time, all of Sun’s products used the SPARC CPU. Unlike many other computer companies, Sun decided not to manufacture the SPARC CPU chip itself. Instead, it licensed several different semiconductor manufacturers to produce them, hoping that competition among them would drive performance up and prices down. These vendors produced a number of different chips, based on different technologies, running at different clock speeds, and with various prices. These chips included the MicroSPARC, HyperSPARC, Super-

SEC. 1.4

EXAMPLE COMPUTER FAMILIES

43

SPARC, and TurboSPARC. Although these CPUs differed in minor ways, all were binary compatible and ran the same user programs without modification. Sun always wanted SPARC to be an open architecture, with many suppliers of parts and systems, in order to build an industry that could compete in a personal computer world already dominated by Intel-based CPUs. To gain the trust of companies that were interested in the SPARC but did not want to invest in a product controlled by a competitor, Sun created an industry consortium, SPARC International, to manage the development of future versions of the SPARC architecture. Thus it is important to distinguish between the SPARC architecture, which is a specification of the instruction set and other programmer-visible features, and a particular implementation of it. In this book we will study both the generic SPARC architecture, and, when discussing CPU chips in Chaps. 3 and 4, a specific SPARC chip used in Sun workstations. The initial SPARC was a full 32-bit machine, running at 36 MHz. The CPU, called the IU (Integer Unit) was lean and mean, with only three major instruction formats and only 55 instructions in all. In addition, a floating-point unit added another 14 instructions. This history can be contrasted to the Intel line, which started out with 8- and 16-bit chips (8088, 8086, 80286) and finally became a 32bit chip with the 80386. The SPARC’s first break with the past occurred in 1995, with the development of Version 9 of the SPARC architecture, a full 64-bit architecture, with 64bit addresses and 64-bit registers. The first Sun workstation to implement the V9 (Version 9) architecture was the UltraSPARC I, introduced in 1995 (Tremblay and O’Connor, 1996). Despite its being a 64-bit machine, it was also fully binary compatible with the existing 32-bit SPARCs. The UltraSPARC was intended to break new ground. Whereas previous machines were designed for handling alphanumeric data and running programs like word processors and spreadsheets, the UltraSPARC was designed from the beginning to handle images, audio, video, and multimedia in general. Among other innovations besides the 64-bit architecture were 23 new instructions, including some for packing and unpacking pixels from 64-bit words, scaling and rotating images, block moves, and performing real-time video compression and decompression. These instructions, called VIS (Visual Instruction Set) were aimed at providing general multimedia capability, analogous to Intel’s MMX instructions. The UltraSPARC was aimed at high-end applications, such as large multiprocessor Web servers with dozens of CPUs and physical memories of up to 8 TB [1 TB (terabyte) = 1012 bytes]. However, smaller versions can be used in notebook computers as well. The successors to the UltraSPARC I were the UltraSPARC II, UltraSPARC III, and UltraSPARC IV. These models differ primarily in clock speed, but some new features were added in each iteration as well. In most of this book, when we discuss the SPARC architecture, we will primarily use the 64-bit V9 UltraSPARC III Cu as our example. The UltraSPARC IV is essentially a dual processor in

44

INTRODUCTION

CHAP. 1

which two UltraSPARC IIIs are colocated on the same CPU chip and share the same memory. We will cover it when we come to multiprocessors in Chap. 8.

1.4.3 Introduction to the 8051 Our third example is very different from our first (the Pentium 4, used in personal computers) and the second (the UltraSPARC III, used in servers). It is the 8051, which is used in embedded systems. The 8051 story starts in 1976, when the 8-bit 8080 had been on the market for about two years. Appliance makers were starting to incorporate the 8080 into their devices, but to build a complete system they needed the 8080 CPU chip, one or more memory chips, and one or more I/O chips. The cost of at least three chips and their interconnection was substantial, and restricted the use of computers in embedded systems to fairly large and expensive items. Many manufacturers asked Intel to put the whole computer (CPU, memory, and I/O) on a single chip to reduce costs. Intel responded to its customers by producing the 8748 chip, a 17,000transistor microcontroller containing an 8080-like CPU, 1 KB of read-only memory for the program, 64 bytes of read/write memory for the variables, an 8-bit timer, and 27 I/O lines for controlling switches, buttons, and lights. While primitive, the chip was a commercial success, which led Intel to release the 8051 in 1980. This new chip contained 60,000 transistors, a much faster CPU, 4 KB of read-only memory, 128 bytes of read-write memory, 32 I/O lines, a serial port, and two 16-bit timers. It was soon followed by other members of what Intel called the MCS-51 family, shown in Fig. 1-14. ################################################################## ! Chip ! Program memory ! Mem. type ! RAM ! Timers ! Interrupts ! ################################################################## ! ! ! ! ! ! ! 8031 ! 0 KB ! ! 5 ! 128 ! 2 ! !################################################################## ! ! ! ! ! ! 8051 ! 4 KB ROM 128 2 5 ! ! ! ! ! ! !################################################################## ! ! 5 ! 128 ! 2 ! EPROM !################################################################## 8751 ! 8 KB ! ! ! ! ! ! ! 0 KB ################################################################## ! ! 6 ! 256 ! 3 ! ! 8032 ! ! ! 6 ! 256 ! 3 ! ROM !################################################################## 8052 ! 8 KB ! ! ! ! ! ! ! 8752 ! 8 KB ! ! 6 ! 256 ! 3 ! EPROM !################################################################## Figure 1-14. Members of the MCS-51 family.

All of these chips use read-only memories for the program plus a small amount of read-write memory, called RAM (Random Access Memory) for data storage. With the 8031 and 8032, the program memory is external, allowing more than 8 KB to be used if needed. We will study ROM (Read Only Memory) and EPROM (Erasable Programmable ROM) in Chap. 3. For the moment, it is sufficient to know that the 8051 and 8052 are single chip microcontrollers used in actual products. Each batch is custom manufactured for the customer (e.g., an appliance manufacturer) and contains the program supplied by the customer.

SEC. 1.4

EXAMPLE COMPUTER FAMILIES

45

In order to develop the software, however, the customer needs a development system. That is where the 8751 and 8752 come in. They are much more expensive than the 8051 and 8052, but can be programmed by the customer for the purpose of software testing. If a bug is found in the code, the program in the 8751 or 8752 can be erased by exposing the chip to an ultraviolet light. A new program can then be burned into it. When the software is finished, it can be delivered to the chip manufacturer, which then produces custom 8051s or 8052s containing the code. In terms of architecture, interfacing, and programming, all the MCS-51 family members are very similar. For simplicity, we will mostly refer to the 8051, pointing out differences with the other chips where that is needed. To some people, using an 8-bit chip that is more than 20 years old as an example may seem like a strange idea, but there are good reasons for doing so. The number of microcontrollers sold each year is around 8 billion, and climbing rapidly. This number is orders of magnitude more than the number of Pentiums sold annually. It was not until 2001 until the number of 8-bit microcontrollers sold per year exceeded the volume of the 4-bit microcontrollers. Currently 8-bit microcontrollers outsell all the others combined, and the MCS-51 family is the most popular 8-bit family. Given the growing importance of embedded systems, anyone studying computer architecture should be familiar with the chips used in them, and the 8051 is one of the most popular ones. There are a variety of reasons for the success of the 8051. First, and foremost, is the price. Depending on the number of units ordered, an 8051 can be had for around 10 to 15 cents per chip, maybe less in large volume. In contrast, a 32-bit microcontroller typically costs 30 times as much, with a 16-bit one somewhere in between. For products that sell for under $50 in competitive markets, knocking a couple of dollars off the manufacturing cost can make a significant difference in retail price and in sales. This is the main reason the 8051 is so popular—it is very cheap. Second, over half a dozen companies manufacture 8051s under license from Intel. Their products cover a wide range of speeds, from the original 12 MHz to 100 MHz, and use many different manufacturing and packaging technologies. Not only does this competition keep prices low, but large customers are much happier when they are not dependent on a single supplier. Third, because the 8051 has been around for so long, there is a vast amount of software for it, including assemblers, compilers for C and other languages, libraries of all kinds, debuggers, simulators, test software, and much more. There are also many complete development systems on the market, which speed up developing the embedded hardware and software. Finally, large numbers of programmers and hardware engineers are familiar with the 8051, making it easy to find skilled personnel. This popularity feeds on itself. Researchers interested in embedded systems often choose the 8051 as their main object of study due to its widespread use, for

46

INTRODUCTION

CHAP. 1

example, for testing out new energy-efficient technologies (Martin et al., 2003) or fault tolerance (Lima et al., 2002). There is a great deal of information about the 8051 on the Internet. One good starting place is www.8051.com. In addition, people are still writing new books about it (Ayala, 2004; Calcutt et al., 2004; MacKenzie et al., 2005; and Mazidi et al., 2005).

1.5 METRIC UNITS To avoid any confusion, it is worth stating explicitly that in this book, as in computer science in general, metric units are used instead of traditional English units (the furlong-stone-fortnight system). The principal metric prefixes are listed in Fig. 1-15. The prefixes are typically abbreviated by their first letters, with the units greater than 1 capitalized (KB, MB, etc.). One exception (for historical reasons) is kbps for kilobits/sec. Thus, a 1-Mbps communication line transmits 106 bits/sec and a 100 psec (or 100 ps) clock ticks every 10−10 seconds. Since milli and micro both begin with the letter ‘‘m,’’ a choice had to be made. Normally, ‘‘m’’ is for milli and ‘‘µ’’ (the Greek letter mu) is for micro. # ################################################################################################# ! ! ! Prefix ! Exp. ! ! Prefix ! Explicit Explicit #! Exp. ################################################################################################# ! ! ! ! ! ! −3 ! 0.001 ! milli ! 103 ! 1,000 ! Kilo ! #!! 10 ################################################################################################# ! ! ! ! ! ! −6 1,000,000 ! Mega ! ! 0.000001 ! micro ! 106 ! #! 10 ################################################################################################# ! ! ! ! ! ! ! −9 9 1,000,000,000 ! Giga ! #! 10 ################################################################################################# ! 0.000000001 ! nano ! 10 ! −12 ! ! ! pico ! 1012 ! 0.000000000001 1,000,000,000,000 ! Tera ! #! 10 ################################################################################################# ! ! ! ! ! ! −15 ! 0.000000000000001 ! femto ! 1015 ! 1,000,000,000,000,000 ! Peta ! #!! 10 ################################################################################################# ! ! ! ! ! ! −18 1,000,000,000,000,000,000 ! Exa ! !#10 ! 0.0000000000000000001 ! atto ! 1018 ! ################################################################################################# ! ! ! ! ! ! −21 ! 21 1,000,000,000,000,000,000,000 ! Zetta ! !#10 ################################################################################################# ! 0.0000000000000000000001 ! zepto ! 10 ! ! 10−24 ! 0.0000000000000000000000001 ! yocto ! 1024 ! 1,000,000,000,000,000,000,000,000 ! Yotta ! !# ################################################################################################# ! ! ! ! ! !

Figure 1-15. The principal metric prefixes.

It is also worth pointing out that for measuring memory, disk, file, and database sizes, in common industry practice, the units have slightly different meanings. There, kilo means 210 (1024) rather than 103 (1000) because memories are always a power of two. Thus, a 1-KB memory contains 1024 bytes, not 1000 bytes. Similarly, a 1-MB memory contains 220 (1,048,576) bytes, a 1-GB memory contains 230 (1,073,741,824) bytes, and a 1-TB database contains 240 (1,099,511,627,776) bytes. However, a 1-kbps communication line can transmit 1000 bits per second and a 10-Mbps LAN runs at 10,000,000 bits/sec because these speeds are not powers of two. Unfortunately, many people tend to mix up these two systems, especially for disk sizes. To avoid ambiguity, in this book, we

SEC. 1.5

METRIC UNITS

47

will use the symbols KB, MB, GB, and TB for 210 , 220 , 230 , and 240 bytes, respectively, and the symbols kbps, Mbps, Gbps, and Tbps for 103 , 106 , 109 , and 1012 bits/sec, respectively.

1.6 OUTLINE OF THIS BOOK This book is about multilevel computers (which includes nearly all modern computers) and how they are organized. We will examine four levels in considerable detail—namely, the digital logic level, the microarchitecture level, the ISA level, and the operating system machine level. Some of the basic issues to be examined include the overall design of the level (and why it was designed that way), the kinds of instructions and data available, the memory organization and addressing, and the method by which the level is implemented. The study of these topics, and similar ones, is called computer organization or computer architecture. We are primarily concerned with concepts rather than details or formal mathematics. For that reason, some of the examples given will be highly simplified, in order to emphasize the central ideas and not the details. To provide some insight into how the principles presented in this book can be, and are, applied in practice, we will use the Pentium 4, the UltraSPARC III, and the 8051 as running examples throughout the book. These three have been chosen for several reasons. First, all are widely used and the reader is likely to have access to at least one of them. Second, each one has its own unique architecture, which provides a basis for comparison and encourages a ‘‘what are the alternatives?’’ attitude. Books dealing with only one machine often leave the reader with a ‘‘true machine design revealed’’ feeling, which is absurd in light of the many compromises and arbitrary decisions that designers are forced to make. The reader is encouraged to study these and all other computers with a critical eye and to try to understand why things are the way they are, as well as how they could have been done differently rather than simply accepting them as given. It should be made clear from the beginning that this is not a book about how to program the Pentium 4, UltraSPARC III, or 8051. These machines will be used for illustrative purposes where appropriate, but we make no pretense of being complete. Readers wishing a thorough introduction to one of them should consult the manufacturer’s publications. Chapter 2 is an introduction to the basic components of a computer— processors, memories, and input/output equipment. It is intended to provide an overview of the system architecture and an introduction to the following chapters. Chapters 3, 4, 5, and 6 each deal with one specific level shown in Figure 1-2. Our treatment is bottom-up, because machines have traditionally been designed that way. The design of level k is largely determined by the properties of level k − 1, so it is hard to understand any level unless you already have a good grasp of the underlying level that motivated it. Also, it is educationally sound to

48

INTRODUCTION

CHAP. 1

proceed from the simpler lower levels to the more complex higher levels rather than vice versa. Chapter 3 is about the digital logic level, the machine’s true hardware. It discusses what gates are and how they can be combined into useful circuits. Boolean algebra, a tool for analyzing digital circuits, is also introduced. Computer buses are explained, especially the popular PCI bus. Numerous examples from industry are discussed in this chapter, including the three running examples mentioned above. Chapter 4 introduces the architecture of the microarchitecture level and its control. Since the function of this level is to interpret the level 2 instructions in the layer above it, we will concentrate on this topic and illustrate it by means of examples. The chapter also contains discussions of the microarchitecture level of some real machines. Chapter 5 discusses the ISA level, the one most computer vendors advertise as the machine language. We will look at our example machines here in detail. Chapter 6 covers some of the instructions, memory organization, and control mechanisms present at the operating system machine level. The examples used here are Windows XP (popular on high-end Pentium 4 server systems) and UNIX, used on the UltraSPARC III. Chapter 7 is about the assembly language level. It covers both assembly language and the assembly process. The subject of linking also comes up here. Chapter 8 discusses parallel computers, an increasingly important topic nowadays. Some of these parallel computers have multiple CPUs that share a common memory. Others have multiple CPUs without common memory. Some are supercomputers; some are systems on a chip; others are COWs. Chapter 9 contains an annotated list of suggested readings, arranged by subject, and an alphabetical list of literature citations. It is the most important chapter in the book. Use it.

PROBLEMS 1. Explain each of the following terms in your own words: a. Translator. b. Interpreter. c. Virtual machine. 2. What is the difference between interpretation and translation? 3. Is it conceivable for a compiler to generate output for the microarchitecture level instead of for the ISA level? Discuss the pros and cons of this proposal. 4. Can you imagine any multilevel computer in which the device level and digital logic levels were not the lowest levels? Explain.

CHAP. 1

PROBLEMS

49

5. Consider a multilevel computer in which all the levels are different. Each level has instructions that are m times as powerful as those of the level below it; that is, one level r instruction can do the work of m level r − 1 instructions. If a level 1 program requires k seconds to run, how long would equivalent programs take at levels 2, 3, and 4, assuming n level r instructions are required to interpret a single r + 1 instruction? 6. Some instructions at the operating system machine level are identical to ISA language instructions. These instructions are carried out directly by the microprogram rather than by the operating system. In light of your answer to the preceding problem, why do you think this is the case? 7. Consider a computer with identical interpreters at levels 1, 2, and 3. It takes an interpreter n instructions to fetch, examine, and execute one instruction. A level 1 instruction takes k nanoseconds to execute. How long does it take for an instruction at levels 2, 3, and 4? 8. In what sense are hardware and software equivalent? Not equivalent? 9. Babbage’s difference engine had a fixed program that could not be changed. Is this essentially the same thing as a modern CD-ROM that cannot be changed? Explain your answer. 10. One of the consequences of von Neumann’s idea to store the program in memory is that programs can be modified, just like data. Can you think of an example where this facility might have been useful? (Hint: Think about doing arithmetic on arrays.) 11. The performance ratio of the 360 model 75 was 50 times that of the 360 model 30, yet the cycle time was only five times as fast. How do you account for this discrepancy? 12. Two basic system designs are shown in Figure 1-5 and Figure 1-6. Describe how input/output might occur in each system. Which one has the potential for better overall system performance? 13. Suppose that each of the 300 million people in the United States fully consumes two packages of goods a day bearing RFID tags. How many RFID tags have to be produced annually to meet that demand? At a penny a tag, what is the total cost of the tags? Given the size of GDP, is this amount of money going to be an obstacle to their use on every package offered for sale? 14. Name three appliances that are candidates for being run by an embedded CPU. 15. At a certain point in time, a transistor on a microprocessor was 0.1 micron in diameter. According to Moore’s law, how big would a transistor be on next year’s model? 16. The legal issue of who invented the computer was settled in April 1973 by Judge Earl Larson, who handled a patent infringement lawsuit filed by the Sperry Rand Corporation, which had acquired the ENIAC patents. Sperry Rand’s position was that everybody making a computer owed them royalties because it owned the key patents. The case went to trial in June 1971 and over 30,000 exhibits were entered. The court transcript ran to over 20,000 pages. Study this case more carefully using the extensive information available on the Internet and write a report discussing the technical aspects of the case. What exactly did Eckert and Mauchley patent and why did the judge feel their system was based on Atanasoff’s earlier work?

50

INTRODUCTION

CHAP. 1

17. Pick the three people you think were most influential in creating modern computer hardware and write a short report describing their contributions and why you picked them. 18. Repeat the previous question for computer software.

2 COMPUTER SYSTEMS ORGANIZATION

A digital computer consists of an interconnected system of processors, memories, and input/output devices. This chapter is an introduction to these three components and to their interconnection, as background for the detailed examination of specific levels in the five succeeding chapters. Processors, memories, and input/output are key concepts and will be present at every level, so we will start our study of computer architecture by looking at all three in turn.

2.1 PROCESSORS The organization of a simple bus-oriented computer is shown in Fig. 2-1. The CPU (Central Processing Unit) is the ‘‘brain’’ of the computer. Its function is to execute programs stored in the main memory by fetching their instructions, examining them, and then executing them one after another. The components are connected by a bus, which is a collection of parallel wires for transmitting address, data, and control signals. Buses can be external to the CPU, connecting it to memory and I/O devices, but also internal to the CPU, as we will see shortly. The CPU is composed of several distinct parts. The control unit is responsible for fetching instructions from main memory and determining their type. The arithmetic logic unit performs operations such as addition and Boolean AND needed to carry out the instructions. 51

52

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Central processing unit (CPU) Control unit Arithmetic logical unit (ALU)

I/O devices

Registers

…

…

Main memory

Disk

Printer

Bus

Figure 2-1. The organization of a simple computer with one CPU and two I/O devices.

The CPU also contains a small, high-speed memory used to store temporary results and certain control information. This memory is made up of a number of registers, each of which has a certain size and function. Usually, all the registers have the same size. Each register can hold one number, up to some maximum determined by the size of the register. Registers can be read and written at high speed since they are internal to the CPU. The most important register is the Program Counter (PC), which points to the next instruction to be fetched for execution. ( The name ‘‘program counter’’ is somewhat misleading because it has nothing to do with counting anything, but the term is universally used. Also important is the Instruction Register (IR), which holds the instruction currently being executed. ( Most computers have numerous other registers as well, some of them general purpose as well as some for specific purposes.

2.1.1 CPU Organization The internal organization of part of a typical von Neumann CPU is shown in Fig. 2-2 in more detail. This part is called the data path and consists of the registers (typically 1 to 32), the ALU (Arithmetic Logic Unit), and several buses connecting the pieces. The registers feed into two ALU input registers, labeled A and B in the figure. These registers hold the ALU input while the ALU is performing

SEC. 2.1

53

PROCESSORS

some computation. The data path is very important in all machines and we will discuss it at great length throughout this book. A+B

A

Registers

B

A

B

ALU input register ALU input bus

ALU

A+B

ALU output register

Figure 2-2. The data path of a typical von Neumann machine.

The ALU itself performs addition, subtraction, and other simple operations on its inputs, thus yielding a result in the output register. This output register can be stored back into a register. Later on, the register can be written (i.e., stored) into memory, if desired. Not all designs have the A, B, and output registers. In the example, addition is illustrated. Most instructions can be divided into one of two categories: register-memory or register-register. Register-memory instructions allow memory words to be fetched into registers, where they can be used as ALU inputs in subsequent instructions, for example. (‘‘Words’’ are the units of data moved between memory and registers. A word might be an integer. We will discuss memory organization later in this chapter.) Other register-memory instructions allow registers to be stored back into memory. The other kind of instruction is register-register. A typical register-register instruction fetches two operands from the registers, brings them to the ALU input registers, performs some operation on them, for example, addition or Boolean

54

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

AND, and stores the result back in one of the registers. The process of running two operands through the ALU and storing the result is called the data path cycle and is the heart of most CPUs. To a considerable extent, it defines what the machine can do. The faster the data path cycle is, the faster the machine runs.

2.1.2 Instruction Execution The CPU executes each instruction in a series of small steps. Roughly speaking, the steps are as follows: 1. Fetch the next instruction from memory into the instruction register. 2. Change the program counter to point to the following instruction. 3. Determine the type of instruction just fetched. 4. If the instruction uses a word in memory, determine where it is. 5. Fetch the word, if needed, into a CPU register. 6. Execute the instruction. 7. Go to step 1 to begin executing the following instruction. This sequence of steps is frequently referred to as the fetch-decode-execute cycle. It is central to the operation of all computers. This description of how a CPU works closely resembles a program written in English. Figure 2-3 shows this informal program rewritten as a Java method (i.e., procedure) called interpret. The machine being interpreted has two registers visible to user programs: the program counter (PC), for keeping track of the address of the next instruction to be fetched, and the accumulator (AC), for accumulating arithmetic results. It also has internal registers for holding the current instruction during its execution (instr), the type of the current instruction (instr"type), the address of the instruction’s operand (data"loc), and the current operand itself (data). Instructions are assumed to contain a single memory address. The memory location addressed contains the operand, for example, the data item to add to the accumulator. The very fact that it is possible to write a program that can imitate the function of a CPU shows that a program need not be executed by a ‘‘hardware’’ CPU consisting of a box full of electronics. Instead, a program can be carried out by having another program fetch, examine, and execute its instructions. A program (such as the one in Fig. 2-3) that fetches, examines, and executes the instructions of another program is called an interpreter, as mentioned in Chap. 1. This equivalence between hardware processors and interpreters has important implications for computer organization and the design of computer systems. After having specified the machine language, L, for a new computer, the design team

SEC. 2.1

PROCESSORS

public class Interp { static int PC; static int AC; static int instr; static int instr"type; static int data"loc; static int data; static boolean run"bit = true;

55

// program counter holds address of next instr // the accumulator, a register for doing arithmetic // a holding register for the current instruction // the instruction type (opcode) // the address of the data, or −1 if none // holds the current operand // a bit that can be turned off to halt the machine

public static void interpret(int memory[ ], int starting"address) { // This procedure interprets programs for a simple machine with instructions having // one memory operand. The machine has a register AC (accumulator), used for // arithmetic. The ADD instruction adds an integer in memory to the AC, for example. // The interpreter keeps running until the run bit is turned off by the HALT instruction. // The state of a process running on this machine consists of the memory, the // program counter, the run bit, and the AC. The input parameters consist of // of the memory image and the starting address. PC = starting"address; while (run"bit) { instr = memory[PC]; PC = PC + 1; instr"type = get"instr"type(instr); data"loc = find"data(instr, instr"type); if (data"loc >= 0) data = memory[data "loc]; execute(instr "type, data); }

// fetch next instruction into instr // increment program counter // determine instruction type // locate data (−1 if none) // if data"loc is −1, there is no operand // fetch the data // execute instruction

}

}

private static int get"instr"type(int addr) { ... } private static int find"data(int instr, int type) { ... } private static void execute(int type, int data) { ... }

Figure 2-3. An interpreter for a simple computer (written in Java).

can decide whether they want to build a hardware processor to execute programs in L directly or whether they want to write an interpreter to interpret programs in L instead. If they choose to write an interpreter, they must also provide some hardware machine to run the interpreter. Certain hybrid constructions are also possible, with some hardware execution as well as some software interpretation. An interpreter breaks the instructions of its target machine into small steps. As a consequence, the machine on which the interpreter runs can be much simpler and less expensive than a hardware processor for the target machine would be. This saving is especially significant if the target machine has a large number of instructions and the instructions are fairly complicated, with many options. The

56

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

saving comes essentially from the fact that hardware is being replaced by software (the interpreter) and it costs more to replicate hardware than software. Early computers had small, simple sets of instructions. But the quest for more powerful computers led, among other things, to more powerful individual instructions. Very early on, it was discovered that more complex instructions often led to faster program execution even though individual instructions might take longer to execute. A floating-point instruction is an example of a more complex instruction. Direct support for accessing array elements is another. Sometimes it was as simple as observing that the same two instructions often occurred consecutively, so a single instruction could accomplish the work of both. The more complex instructions were better because the execution of individual operations could sometimes be overlapped or otherwise executed in parallel using different hardware. For expensive, high-performance computers, the cost of this extra hardware could be readily justified. Thus expensive, high-performance computers came to have many more instructions than lower-cost ones. However, the rising cost of software development and instruction compatibility requirements created the need to implement complex instructions even on low-cost computers where cost was more important than speed. By the late 1950s, IBM (then the dominant computer company) had recognized that supporting a single family of machines, all of which executed the same instructions, had many advantages, both for IBM and for its customers. IBM introduced the term architecture to describe this level of compatibility. A new family of computers would have one architecture but many different implementations that could all execute the same program, differing only in price and speed. But how to build a low-cost computer that could execute all the complicated instructions of high-performance, expensive machines? The answer lay in interpretation. This technique, first suggested by Wilkes (1951), permitted the design of simple, lower-cost computers that could nevertheless execute a large number of instructions. The result was the IBM System/360 architecture, a compatible family of computers, spanning nearly two orders of magnitude, both in price and capability. A direct hardware (i.e., not interpreted) implementation was used only on the most expensive models. Simple computers with interpreted instructions also had other benefits. Among the most important were 1. The ability to fix incorrectly implemented instructions in the field, or even make up for design deficiencies in the basic hardware. 2. The opportunity to add new instructions at minimal cost, even after delivery of the machine. 3. Structured design that permitted efficient development, testing, and documenting of complex instructions. As the market for computers exploded dramatically in the 1970s and computing

SEC. 2.1

PROCESSORS

57

capabilities grew rapidly, the demand for low-cost computers favored designs of computers using interpreters. The ability to tailor the hardware and the interpreter for a particular set of instructions emerged as a highly cost-effective design for processors. As the underlying semiconductor technology advanced rapidly, the advantages of the cost outweighed the opportunities for higher performance, and interpreter-based architectures became the conventional way to design computers. Nearly all new computers designed in the 1970s, from minicomputers to mainframes, were based on interpretation. By the late 70s, the use of simple processors running interpreters had become very widespread except among the most expensive, highest-performance models, such as the Cray-1 and the Control Data Cyber series. The use of an interpreter eliminated the inherent cost limitations of complex instructions, and architectures began to explore much more complex instructions, particularly the ways to specify the operands to be used. This trend reached its zenith with Digital Equipment Corporation’s VAX computer, which had several hundred instructions, and more than 200 different ways of specifying the operands to be used in each instruction. Unfortunately, the VAX architecture was conceived from the beginning to be implemented with an interpreter, with little thought given to the implementation of a high-performance model. This mind set resulted in the inclusion of a very large number of instructions of marginal value and which were difficult to execute directly. This omission proved to be fatal to the VAX, and ultimately to DEC as well (Compaq bought DEC in 1998 and Hewlett-Packard bought Compaq in 2001). Though the earliest 8-bit microprocessors were very simple machines with very simple instruction sets, by the late 70s, even microprocessors had switched to interpreter-based designs. During this period, one of the major challenges facing microprocessor designers was dealing with the growing complexity possible through integrated circuits. A major advantage of the interpreter-based approach was the ability to design a simple processor, with the complexity largely confined to the memory holding the interpreter. Thus a complex hardware design could be turned into a complex software design. The success of the Motorola 68000, which had a large interpreted instruction set, and the concurrent failure of the Zilog Z8000 (which had an equally large instruction set, but without an interpreter) demonstrated the advantages of an interpreter for bringing a new microprocessor to market quickly. This success was all the more surprising given Zilog’s head start (the Z8000’s predecessor, the Z80, was far more popular than the 68000’s predecessor, the 6800). Of course, other factors were instrumental here too, not the least of which was Motorola’s long history as a chip manufacturer and Exxon’s (Zilog’s owner) long history of being an oil company, not a chip manufacturer. Another factor working in favor of interpretation during that era was the existence of fast read-only memories, called control stores, to hold the interpreters. Suppose that a typical interpreted instruction took the interpreter 10

58

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

instructions, called microinstructions, at 100 nsec each, and two references to main memory, at 500 nsec each. Total execution time was then 2000 nsec, only a factor of two worse than the best that direct execution could achieve. Had the control store not been available, the instruction would have taken 6000 nsec. A factor of six penalty is a lot harder to swallow than a factor of two penalty.

2.1.3 RISC versus CISC During the late 70s there was experimentation with very complex instructions, made possible by the interpreter. Designers tried to close the ‘‘semantic gap’’ between what machines could do and what high-level programming languages required. Hardly anyone thought about designing simpler machines, just as now not a lot of research goes into designing less powerful operating systems, networks, word processors, etc. (perhaps unfortunately). One group that bucked the trend and tried to incorporate some of Seymour Cray’s ideas in a high-performance minicomputer was led by John Cocke at IBM. This work led to an experimental minicomputer, named the 801, Although IBM never marketed this machine and the results were not published until years later (Radin, 1982), word got out and other people began investigating similar architectures. In 1980, a group at Berkeley led by David Patterson and Carlo Se´quin began designing VLSI CPU chips that did not use interpretation (Patterson, 1985; Patterson and Se´quin, 1982). They coined the term RISC for this concept and named their CPU chip the RISC I CPU followed shortly by the RISC II. Slightly later, in 1981, across the San Francisco Bay at Stanford, John Hennessy designed and fabricated a somewhat different chip he called the MIPS (Hennessy, 1984). These chips evolved into commercially important products, the SPARC and the MIPS, respectively. These new processors were significantly different than commercial processors of the day. Since these new CPUs did not have to be backward compatible with existing products, their designers were free to choose new instruction sets that would maximize total system performance. While the initial emphasis was on simple instructions that could be executed quickly, it was soon realized that designing instructions that could be issued (started) quickly was the key to good performance. How long an instruction actually took mattered less than how many could be started per second. At the time these simple processors were being first designed, the characteristic that caught everyone’s attention was the relatively small number of instructions available, typically around 50. This number was far smaller than the 200 to 300 on established computers such as the DEC VAX and the large IBM mainframes. In fact, the acronym RISC stands for Reduced Instruction Set Computer, which was contrasted with CISC, which stands for Complex Instruction Set Computer (a thinly-veiled reference to the VAX, which dominated university

SEC. 2.1

PROCESSORS

59

Computer Science Departments at the time). Nowadays, few people think that the size of the instruction set is a major issue, but the name stuck. To make a long story short, a great religious war ensued, with the RISC supporters attacking the established order (VAX, Intel, large IBM mainframes). They claimed that the best way to design a computer was to have a small number of simple instructions that execute in one cycle of the data path of Figure 2-2, namely, fetching two registers, combining them somehow (e.g., adding or ANDing them), and storing the result back in a register. Their argument was that even if a RISC machine takes four or five instructions to do what a CISC machine does in one instruction, if the RISC instructions are 10 times as fast (because they are not interpreted), RISC wins. It is also worth pointing out that by this time the speed of main memories had caught up to the speed of read-only control stores, so the interpretation penalty had greatly increased, strongly favoring RISC machines. One might think that given the performance advantages of RISC technology, RISC machines (such as the Sun UltraSPARC) would have mowed over CISC machines (such as the Intel Pentium) in the marketplace. Nothing like this has happened. Why not? First of all, there is the issue of backward compatibility and the billions of dollars companies have invested in software for the Intel line. Second, surprisingly, Intel has been able to employ the same ideas even in a CISC architecture. Starting with the 486, the Intel CPUs contain a RISC core that executes the simplest (and typically most common) instructions in a single data path cycle, while interpreting the more complicated instructions in the usual CISC way. The net result is that common instructions are fast and less common instructions are slow. While this hybrid approach is not as fast as a pure RISC design, it gives competitive overall performance while still allowing old software to run unmodified.

2.1.4 Design Principles for Modern Computers Now that more than two decades have passed since the first RISC machines were introduced, certain design principles have come to be accepted as a good way to design computers given the current state of the hardware technology. If a major change in technology occurs (e.g., a new manufacturing process suddenly makes memory cycle time 10 times faster than CPU cycle time), all bets are off. Thus machine designers should always keep an eye out for technological changes that may affect the balance among the components. That said, there is a set of design principles, sometimes called the RISC design principles, that architects of general-purpose CPUs do their best to follow. External constraints, such as the requirement of being backward compatible with some existing architecture, often require compromises from time to time, but these principles are goals that most designers strive to meet. Below we will discuss the major ones.

60

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

All Instructions Are Directly Executed by Hardware All common instructions are directly executed by the hardware. They are not interpreted by microinstructions. Eliminating a level of interpretation provides high speed for most instructions. For computers that implement CISC instruction sets, the more complex instructions may be broken into separate parts, which can then be executed as a sequence of microinstructions. This extra step slows the machine down, but for less frequently occurring instructions it may be acceptable. Maximize the Rate at Which Instructions Are Issued Modern computers resort to many tricks to maximize their performance, chief among which is trying to start as many instructions per second as possible. After all, if you can issue 500 million instructions/sec, you have built a 500-MIPS processor, no matter how long the instructions actually take to complete. (MIPS stands for Millions of Instructions Per Second; the MIPS processor was so-named as to be a pun on this acronym.) This principle suggests that parallelism can play a major role in improving performance, since issuing large numbers of slow instructions in a short time interval is only possible if multiple instructions can execute at once. Although instructions are always encountered in program order, they are not always issued in program order (because some needed resource might be busy) and they need not finish in program order. Of course, if instruction 1 sets a register and instruction 2 uses that register, great care must be taken to make sure that instruction 2 does not read the register until it contains the correct value. Getting this right requires a lot of bookkeeping but has the potential for performance gains by executing multiple instructions at once. Instructions Should be Easy to Decode A critical limit on the rate of issue of instructions is decoding individual instructions to determine what resources they need. Anything that can aid this process is useful. That includes making instructions regular, fixed length, with a small number of fields. The fewer different formats for instructions, the better. Only Loads and Stores Should Reference Memory One of the simplest ways to break operations into separate steps is to require that operands for most instructions come from—and return to—CPU registers. The operation of moving operands from memory into registers can be performed in separate instructions. Since access to memory can take a long time, and the delay is unpredictable, these instructions can best be overlapped with other in-

SEC. 2.1

PROCESSORS

61

structions if they do nothing but move operands between registers and memory. This observation means that only LOAD and STORE instructions should reference memory. All other instructions should operate only on registers. Provide Plenty of Registers Since accessing memory is relatively slow, many registers (at least 32) need to be provided, so that once a word is fetched, it can be kept in a register until it is no longer needed. Running out of registers and having to flush them back to memory only to later reload them is undesirable and should be avoided as much as possible. The best way to accomplish this is to have enough registers.

2.1.5 Instruction-Level Parallelism Computer architects are constantly striving to improve performance of the machines they design. Making the chips run faster by increasing their clock speed is one way, but for every new design, there is a limit to what is possible by brute force at that moment in history. Consequently, most computer architects look to parallelism (doing two or more things at once) as a way to get even more performance for a given clock speed. Parallelism comes in two general forms, namely, instruction-level parallelism and processor-level parallelism. In the former, parallelism is exploited within individual instructions to get more instructions/sec out of the machine. In the latter, multiple CPUs work together on the same problem. Each approach has its own merits. In this section we will look at instruction-level parallelism; in the one after it, we will look at processor-level parallelism. Pipelining It has been known for years that the actual fetching of instructions from memory is a major bottleneck in instruction execution speed. To alleviate this problem, computers going back at least as far as the IBM Stretch (1959) have had the ability to fetch instructions from memory in advance, so they would be there when they were needed. These instructions were stored in a set of registers called the prefetch buffer. This way, when an instruction was needed, it could usually be taken from the prefetch buffer rather than waiting for a memory read to complete. In effect, prefetching divides instruction execution up into two parts: fetching and actual execution. The concept of a pipeline carries this strategy much further. Instead of dividing instruction execution into only two parts, it is often divided into many (often a dozen or more) parts, each one handled by a dedicated piece of hardware, all of which can run in parallel.

62

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Figure 2-4(a) illustrates a pipeline with five units, also called stages. Stage 1 fetches the instruction from memory and places it in a buffer until it is needed. Stage 2 decodes the instruction, determining its type and what operands it needs. Stage 3 locates and fetches the operands, either from registers or from memory. Stage 4 actually does the work of carrying out the instruction, typically by running the operands through the data path of Figure 2-2. Finally, stage 5 writes the result back to the proper register. S1

S2

S3

S4

S5

Instruction fetch unit

Instruction decode unit

Operand fetch unit

Instruction execution unit

Write back unit

(a) S1:

1

S2:

2

3

4

5

6

7

8

9

1

2

3

4

5

6

7

8

1

2

3

4

5

6

7

1

2

3

4

5

6

1

2

3

4

5

6

7

8

9

S3: S4: S5: 1

2

3

4 5 Time (b)

…

Figure 2-4. (a) A five-stage pipeline. (b) The state of each stage as a function of time. Nine clock cycles are illustrated.

In Fig. 2-4(b) we see how the pipeline operates as a function of time. During clock cycle 1, stage S1 is working on instruction 1, fetching it from memory. During cycle 2, stage S2 decodes instruction 1, while stage S1 fetches instruction 2. During cycle 3, stage S3 fetches the operands for instruction 1, stage S2 decodes instruction 2, and stage S1 fetches the third instruction. During cycle 4, stage S4 executes instruction 1, S3 fetches the operands for instruction 2, S2 decodes instruction 3, and S1 fetches instruction 4. Finally, during cycle 5, S5 writes the result of instruction 1 back, while the other stages work on the following instructions. Let us consider an analogy to make the concept of pipelining clearer. Imagine a cake factory in which the baking of the cakes and the packaging of the cakes for shipment are separated. Suppose that the shipping department has a long conveyor belt with five workers (processing units) lined up along it. Every 10 sec (the clock cycle), worker 1 places an empty cake box on the belt. The box is carried down to worker 2, who places a cake in it. A little later, the box arrives at worker 3’s station, where it is closed and sealed. Then it continues to worker 4,

SEC. 2.1

PROCESSORS

63

who puts a label on the box. Finally, worker 5 removes the box from the belt and puts it in a large container for later shipment to a supermarket. Basically, this is the way computer pipelining works too: each instruction (cake) goes through several processing steps before emerging completed at the far end. Getting back to our pipeline of Fig. 2-4, suppose that the cycle time of this machine is 2 nsec. Then it takes 10 nsec for an instruction to progress all the way through the five-stage pipeline. At first glance, with an instruction taking 10 nsec, it might appear that the machine can run at 100 MIPS, but in fact it does much better than this. At every clock cycle (2 nsec), one new instruction is completed, so the actual rate of processing is 500 MIPS, not 100 MIPS. Pipelining allows a trade-off between latency (how long it takes to execute an instruction), and processor bandwidth (how many MIPS the CPU has). With a cycle time of T nsec, and n stages in the pipeline, the latency is nT nsec because each instruction passes through n stages, each of which takes T nsec. Since one instruction completes every clock cycle and there are 109 /T clock cycles/second, the number of instructions executed per second is 109 /T. For example, if T = 2 nsec, 500 million instructions are executed each seconds. To get the number of MIPS, we have to divide the instruction execution rate by 1 million to get (109 /T)/106 = 1000/T MIPS. Theoretically, we could measure instruction execution rate in BIPS instead of MIPS, but nobody does that so we will not either. Superscalar Architectures If one pipeline is good, then surely two pipelines are better. One possible design for a dual pipeline CPU, based on Figure 2-4, is shown in Fig. 2-5. Here a single instruction fetch unit fetches pairs of instructions together and puts each one into its own pipeline, complete with its own ALU for parallel operation. To be able to run in parallel, the two instructions must not conflict over resource usage (e.g., registers), and neither must depend on the result of the other. As with a single pipeline, either the compiler must guarantee this situation to hold (i.e., the hardware does not check and gives incorrect results if the instructions are not compatible), or conflicts are detected and eliminated during execution using extra hardware. Although pipelines, single or double, are mostly used on RISC machines (the 386 and its predecessors did not have any), starting with the 486 Intel began introducing data pipelines into its CPUs. The 486 had one pipeline and the original Pentium had two five-stage pipelines roughly as in Fig. 2-5, although the exact division of work between stages 2 and 3 (called decode-1 and decode-2) was slightly different than in our example. The main pipeline, called the u pipeline, could execute an arbitrary Pentium instruction. The second pipeline, called the v pipeline, could execute only simple integer instructions (and also one simple floating-point instruction— FXCH).

64

COMPUTER SYSTEMS ORGANIZATION S1

Instruction fetch unit

CHAP. 2

S2

S3

S4

S5

Instruction decode unit

Operand fetch unit

Instruction execution unit

Write back unit

Instruction decode unit

Operand fetch unit

Instruction execution unit

Write back unit

Figure 2-5. Dual five-stage pipelines with a common instruction fetch unit.

Fixed rules determined whether a pair of instructions were compatible so they could be executed in parallel. If the instructions in a pair were not simple enough or incompatible, only the first one was executed (in the u pipeline). The second one was then held and paired with the instruction following it. Instructions were always executed in order. Thus Pentium-specific compilers that produced compatible pairs could produce faster-running programs than older compilers. Measurements showed that a Pentium running code optimized for it was exactly twice as fast on integer programs as a 486 running at the same clock rate (Pountain, 1993). This gain could be attributed entirely to the second pipeline. Going to four pipelines is conceivable, but doing so duplicates too much hardware (computer scientists, unlike folklore specialists, do not believe in the number three). Instead, a different approach is used on high-end CPUs. The basic idea is to have just a single pipeline but give it multiple functional units, as shown in Fig. 2-6. For example, the Pentium II has a structure similar to this figure. It will be discussed in Chap. 4. The term superscalar architecture was coined for this approach in 1987 (Agerwala and Cocke, 1987). Its roots, however, go back more than 40 years to the CDC 6600 computer. The 6600 fetched an instruction every 100 nsec and passed it off to one of 10 functional units for parallel execution while the CPU went off to get the next instruction. The definition of ‘‘superscalar’’ has evolved somewhat over time. It is now used to describe processors that issue multiple instructions—often four or six—in a single clock cycle. Of course, a superscalar CPU must have multiple functional units to hand all these instructions to. Since superscalar processors generally have one pipeline, they tend to look like Fig. 2-6. Using this definition, the 6600 was technically not superscalar because it issued only one instruction per cycle. However, the effect was almost the same: instructions were issued at a much higher rate than they could be executed. The difference between a CPU with a 100 nsec clock that issues one instruction every cycle to a group of functional units and a CPU with a 400 nsec clock that issues four instructions per cycle to the same group of functional units is very small. In both cases, the key idea is that the issue rate is much higher than the execution rate, with the workload being spread across a collection of functional units.

SEC. 2.1

65

PROCESSORS S4 ALU

ALU S1

S2

S3

Instruction fetch unit

Instruction decode unit

Operand fetch unit

S5 LOAD

Write back unit

STORE

Floating point

Figure 2-6. A superscalar processor with five functional units.

Implicit in the idea of a superscalar processor is that the S3 stage can issue instructions considerably faster than the S4 stage is able to execute them. If the S3 stage issued an instruction every 10 nsec and all the functional units could do their work in 10 nsec, no more than one would ever be busy at once, negating the whole idea. In reality, most of the functional units in stage 4 take appreciably longer than one clock cycle to execute, certainly the ones that access memory or do floating-point arithmetic. As can be seen from the figure, it is possible to have multiple ALUs in stage S4.

2.1.6 Processor-Level Parallelism The demand for ever faster computers seems to be insatiable. Astronomers want to simulate what happened in the first microsecond after the big bang, economists want to model the world economy, and teenagers want to play 3D interactive multimedia games over the Internet with their virtual friends. While CPUs keep getting faster, eventually they are going to run into the problems with the speed of light, which is likely to stay at 20 cm/nanosecond in copper wire or optical fiber, no matter how clever Intel’s engineers are. Faster chips also produce more heat, whose dissipation is a problem. Instruction-level parallelism helps a little, but pipelining and superscalar operation rarely win more than a factor of five or ten. To get gains of 50, 100, or more, the only way is to design computers with multiple CPUs, so we will now take a look at how some of these are organized.

66

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Array Computers A substantial number of problems in the physical sciences and engineering involve arrays or otherwise have a highly regular structure. Often the same calculations are performed on many different sets of data at the same time. The regularity and structure of these programs makes them especially easy targets for speedup through parallel execution. There are two methods that have been used to execute large scientific programs quickly. While these two schemes are remarkably similar in most ways, ironically, one of them is thought of as an extension to a single processor, while the other is thought of as a parallel computer. An array processor consists of a large number of identical processors that perform the same sequence of instructions on different sets of data. The world’s first array processor was the University of Illinois ILLIAC IV computer, illustrated in Fig. 2-7 (Bouknight et al., 1972). The original plan was to build a machine consisting of four quadrants, each quadrant having an 8 × 8 square grid of processor/memory elements. A single control unit per quadrant broadcast instructions, which were carried out in lockstep by all the processors, each one using its own data from its own memory (loaded during the initialization phase). This design, clearly very different from a standard Von Neumann machine, is sometimes referred to as a SIMD (Single Instruction-stream Multiple Datastream) processor. Due to a cost overrun by a factor of four, only one quadrant was ever built, but it did achieve a performance of 50 megaflops (million floating-point operations per second). It is said that had the entire machine been completed and had it achieved its original performance goal (1 gigaflop), it would have doubled the computing power of the entire world. Control unit Broadcasts instructions

Processor

8 × 8 Processor/memory grid

Memory

Figure 2-7. An array processor of the ILLIAC IV type.

A vector processor appears to the programmer very much like an array processor. Like an array processor, it is very efficient at executing a sequence of

SEC. 2.1

PROCESSORS

67

operations on pairs of data elements. But unlike an array processor, all of the addition operations are performed in a single, heavily-pipelined adder. The company Seymour Cray founded, Cray Research, produced many vector processors, starting with the Cray-1 back in 1974 and continuing through current models (Cray Research is now part of SGI). Both array processors and vector processors work on arrays of data. Both execute single instructions that, for example, add the elements together pairwise for two vectors. But while the array processor does it by having as many adders as elements in the vector, the vector processor has the concept of a vector register, which consists of a set of conventional registers that can be loaded from memory in a single instruction, which actually loads them from memory serially. Then a vector addition instruction performs the pairwise addition of the elements of two such vectors by feeding them to a pipelined adder from the two vector registers. The result from the adder is another vector, which can either be stored into a vector register, or used directly as an operand for another vector operation. While no array processors are currently in production, the idea is far from dead. The MMX and SSE instructions available on the Pentium 4 use this execution model to speed up multimedia software. In this respect, the Pentium 4 has the ILLIAC IV as one of its ancestors. Multiprocessors The processing elements in an array processor are not independent CPUs, since there is only one control unit shared among all of them. Our first parallel system with multiple full-blown CPUs is the multiprocessor, a system with more than one CPU sharing a common memory, like a group of people in a room sharing a common blackboard. Since each CPU can read or write any part of memory, they must co-ordinate (in software) to avoid getting in each other’s way. When two or more CPUs have the ability to interact closely, as is the case with multiprocessors, they are said to be tightly coupled. Various implementation schemes are possible. The simplest one is to have a single bus with multiple CPUs and one memory all plugged into it. A diagram of such a bus-based multiprocessor is shown in Fig. 2-8(a). It does not take much imagination to realize that with a large number of fast processors constantly trying to access memory over the same bus, conflicts will result. Multiprocessor designers have come up with various schemes to reduce this contention and improve performance. One design, shown in Fig. 2-8(b), gives each processor some local memory of its own, not accessible to the others. This memory can be used for program code and those data items that need not be shared. Access to this private memory does not use the main bus, greatly reducing bus traffic. Other schemes (e.g., caching) are also possible. Multiprocessors have the advantage over other kinds of parallel computers that the programming model of a single shared memory is an easy one to work

68

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Local memories

Shared memory

Shared memory CPU

CPU

CPU

CPU

CPU

CPU

CPU

CPU

Bus

Bus (a)

(b)

Figure 2-8. (a) A single-bus multiprocessor. (b) A multicomputer with local memories.

with. For example, imagine a program looking for cancer cells in a photograph of some tissue taken through a microscope. The digitized photograph could be kept in the common memory, with each processor assigned some region of the photograph to hunt in. Since each processor has access to the entire memory, studying a cell that starts in its assigned region but straddles the boundary into the next region is no problem. Multicomputers Although multiprocessors with a modest number of processors (≤ 256) are relatively easy to build, large ones are surprisingly difficult to construct. The difficulty is in connecting all the processors to the memory. To get around these problems, many designers have simply abandoned the idea of having a shared memory and just build systems consisting of large numbers of interconnected computers, each having its own private memory, but no common memory. These systems are called multicomputers. The CPUs in a multicomputer are sometimes said to be loosely coupled, to contrast them with the tightly-coupled CPUs in a multiprocessor. The CPUs in a multicomputer communicate by sending each other messages, something like e-mail, but much faster. For large systems, having every computer connected to every other computer is impractical, so topologies such as 2D and 3D grids, trees, and rings are used. As a result, messages from one computer to another often must pass through one or more intermediate computers or switches to get from the source to the destination. Nevertheless, message-passing times on the order of a few microseconds can be achieved without much difficulty. Multicomputers with nearly 10,000 CPUs have been built and put into operation.

SEC. 2.1

PROCESSORS

69

Since multiprocessors are easier to program and multicomputers are easier to build, there is much research on designing hybrid systems that combine the good properties of each. Such computers try to present the illusion of shared memory, without going to the expense of actually constructing it. We will go into multiprocessors and multicomputers in detail in Chap. 8.

2.2 PRIMARY MEMORY The memory is that part of the computer where programs and data are stored. Some computer scientists (especially British ones) use the term store or storage rather than memory, although more and more, the term ‘‘storage’’ is used to refer to disk storage. Without a memory from which the processors can read and write information, there would be no stored-program digital computers.

2.2.1 Bits The basic unit of memory is the binary digit, called a bit. A bit may contain a 0 or a 1. It is the simplest possible unit. (A device capable of storing only zeros could hardly form the basis of a memory system; at least two values are needed.) People often say that computers use binary arithmetic because it is ‘‘efficient.’’ What they mean (although they rarely realize it) is that digital information can be stored by distinguishing between different values of some continuous physical quantity, such as voltage or current. The more values that must be distinguished, the less separation between adjacent values, and the less reliable the memory. The binary number system requires only two values to be distinguished. Consequently, it is the most reliable method for encoding digital information. If you are not familiar with binary numbers, see Appendix A. Some computers, such as the large IBM mainframes, are advertised as having decimal as well as binary arithmetic. This trick is accomplished by using 4 bits to store one decimal digit using a code called BCD (Binary Coded Decimal). Four bits provide 16 combinations, used for the 10 digits 0 through 9, with six combinations not used. The number 1944 is shown below encoded in decimal and in pure binary, using 16 bits in each example: decimal: 0001 1001 0100 0100

binary: 0000011110011000

Sixteen bits in the decimal format can store the numbers from 0 to 9999, giving only 10,000 combinations, whereas a 16-bit pure binary number can store 65,536 different combinations. For this reason, people say that binary is more efficient. However, consider what would happen if some brilliant young electrical engineer invented a highly reliable electronic device that could directly store the digits 0 to 9 by dividing the region from 0 to 10 volts into 10 intervals. Four of

70

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

these devices could store any decimal number from 0 to 9999. Four such devices would provide 10,000 combinations. They could also be used to store binary numbers, by only using 0 and 1, in which case, four of them could only store 16 combinations. With such devices, the decimal system is obviously more efficient.

2.2.2 Memory Addresses Memories consist of a number of cells (or locations) each of which can store a piece of information. Each cell has a number, called its address, by which programs can refer to it. If a memory has n cells, they will have addresses 0 to n − 1. All cells in a memory contain the same number of bits. If a cell consists of k bits, it can hold any one of 2k different bit combinations. Figure 2-9 shows three different organizations for a 96-bit memory. Note that adjacent cells have consecutive addresses (by definition). Address

Address

1 Cell

Address

0

0

0

1

1

1

2

2

2

3

3

3

4

4

4

5

5

5

6

6

16 bits

7

7

(c)

8

12 bits

9

(b)

10 11 8 bits (a)

Figure 2-9. Three ways of organizing a 96-bit memory.

Computers that use the binary number system (including octal and hexadecimal notation for binary numbers) express memory addresses as binary numbers. If an address has m bits, the maximum number of cells addressable is 2m . For example, an address used to reference the memory of Fig. 2-9(a) needs at least 4 bits in order to express all the numbers from 0 to 11. A 3-bit address is sufficient for Fig. 2-9(b) and (c), however. The number of bits in the address determines the

SEC. 2.2

PRIMARY MEMORY

71

maximum number of directly addressable cells in the memory and is independent of the number of bits per cell. A memory with 212 cells of 8 bits each and a memory with 212 cells of 64 bits each need 12-bit addresses. The number of bits per cell for some computers that have been sold commercially is listed in Fig. 2-10. ################################## ! Bits/cell ! Computer !################################## ! ! Burroughs B1700 1 !################################## ! ! ! IBM PC ! ! 8 !################################## ! ! !################################## ! ! DEC PDP-8 12 ! ! ! IBM 1130 16 !################################## ! ! !################################## ! ! DEC PDP-15 18 ! ! ! XDS 940 24 !################################## ! ! ! Electrologica X8 ! ! 27 ################################## ! ! ! XDS Sigma 9 32 !################################## ! ! ! Honeywell 6180 ! ! 36 ################################## ! ! ! ! CDC 3600 ! ! 48 ################################## ! ! ! 60 ! CDC Cyber ################################## ! ! !

Figure 2-10. Number of bits per cell for some historically interesting commercial computers.

The significance of the cell is that it is the smallest addressable unit. In recent years, nearly all computer manufacturers have standardized on an 8-bit cell, which is called a byte. Bytes are grouped into words. A computer with a 32-bit word has 4 bytes/word, whereas a computer with a 64-bit word has 8 bytes/word. The significance of a word is that most instructions operate on entire words, for example, adding two words together. Thus a 32-bit machine will have 32-bit registers and instructions for manipulating 32-bit words, whereas a 64-bit machine will have 64-bit registers and instructions for moving, adding, subtracting, and otherwise manipulating 64-bit words.

2.2.3 Byte Ordering The bytes in a word can be numbered from left-to-right or right-to-left. At first it might seem that this choice is unimportant, but as we shall see shortly, it has major implications. Figure 2-11(a) depicts part of the memory of a 32-bit computer whose bytes are numbered from left-to-right, such as the SPARC or the big IBM mainframes. Figure 2-11(b) gives the analogous representation of a 32bit computer using right-to-left numbering, such as the Intel family. The former system, where the numbering begins at the ‘‘big’’ (i.e., high-order) end is called a big endian computer, in contrast to the little endian of Fig. 2-11(b). These terms

72

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

are due to Jonathan Swift, whose Gulliver’s Travels satirized politicians who made war over their dispute about whether eggs should be broken at the big end or the little end. The term was first used in computer architecture in a delightful article by Cohen (1981). Address

Big endian

Address

Little endian

0

0

1

2

3

3

2

1

0

0

4

4

5

6

7

7

6

5

4

4

8

8

9

10

11

11

10

9

8

8

12

12

13

14

15

15

14

13

12

12

Byte

Byte

32-bit word

32-bit word

(a)

(b)

Figure 2-11. (a) Big endian memory. (b) Little endian memory.

It is important to understand that in both the big endian and little endian systems, a 32-bit integer with the numerical value of, say, 6, is represented by the bits 110 in the rightmost (low-order) 3 bits of a word and zeros in the leftmost 29 bits. In the big endian scheme, the 110 bits are in byte 3 (or 7, or 11, etc.), whereas in the little endian scheme they are in byte 0 (or 4, or 8, etc.). In both cases, the word containing this integer has address 0. If computers only stored integers, there would not be any problem. However, many applications require a mixture of integers, character strings, and other data types. Consider, for example, a simple personnel record consisting of a string (employee name), and two integers (age and department number). The string is terminated with 1 or more 0 bytes to fill out a word. The big endian representation is shown in Fig. 2-12(a); the little endian representation is shown in Fig. 212(b) for Jim Smith, age 21, department 260 (1 × 256 + 4 = 260). Both of these representations are fine and internally consistent. The problems begin when one of the machines tries to send the record to the other one over a network. Let us assume that the big endian sends the record to the little endian one byte at a time, starting with byte 0 and ending with byte 19. (We will be optimistic and assume the bits of the bytes are not reversed by the transmission, as we have enough problems as is.) Thus the big endian’s byte 0 goes into the little endian’s memory at byte 0, and so on, as shown in Fig. 2-12(c). When the little endian tries to print the name, it works fine, but the age comes out as 21 × 224 and the department is just as garbled. This situation arises because the transmission has reversed the order of the characters in a word, as it should, but it has also reversed the bytes in an integer, which it should not. An obvious solution is to have the software reverse the bytes within a word after the copy has been made. Doing this leads to Fig. 2-12(d) which makes the

SEC. 2.2

73

PRIMARY MEMORY

Big endian

Transfer from big endian to little endian

Little endian

0

J

I

M

M

I

J

0

4

S

M

I

T

T

I

M

S

4

8

H

0

0

0

0

0

0

H

8

12

0

16

0

0

0 21

0

0

0 21

0

1

0

0

1

(a)

4

4

(b)

Transfer and swap

M

I

J

J

I

M

T

I

M

S

S

M

I

T

4

0

0

0

H

H

0

0

0

8

12

21 0

0

0

0

0

0 21 12

16

4

0

0

0

0

1

1 (c)

0

4 16

(d)

Figure 2-12. (a) A personnel record for a big endian machine. (b) The same record for a little endian machine. (c) The result of transferring the record from a big endian to a little endian. (d) The result of byte-swapping (c).

two integers fine but turns the string into ‘‘MIJTIMS’’ with the ‘‘H’’ hanging in the middle of nowhere. This reversal of the string occurs because when reading it, the computer first reads byte 0 (a space), then byte 1 (M), and so on. There is no simple solution. One way that works, but is inefficient, is to include a header in front of each data item telling what kind of data follows (string, integer, or other) and how long it is. This allows the receiver to perform the necessary conversions only. In any event, it should be clear that the lack of a standard for byte ordering is a major nuisance when exchanging data between different machines.

2.2.4 Error-Correcting Codes Computer memories can make errors occasionally due to voltage spikes on the power line or other causes. To guard against such errors, some memories use error-detecting or error-correcting codes. When these codes are used, extra bits are added to each memory word in a special way. When a word is read out of memory, the extra bits are checked to see if an error has occurred. To understand how errors can be handled, it is necessary to look closely at what an error really is. Suppose that a memory word consists of m data bits to which we will add r redundant, or check bits. Let the total length be n (i.e., n = m + r). An n-bit unit containing m data and r check bits is often referred to as an n-bit codeword. Given any two codewords, say, 10001001 and 10110001, it is possible to determine how many corresponding bits differ. In this case, 3 bits differ. To determine how many bits differ, just compute the bitwise Boolean EXCLUSIVE OR of the two codewords, and count the number of 1 bits in the result. The number of bit positions in which two codewords differ is called the Hamming distance (Hamming, 1950). Its main significance is that if two codewords are a

74

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Hamming distance d apart, it will require d single-bit errors to convert one into the other. For example, the codewords 11110001 and 00110000 are a Hamming distance 3 apart because it takes 3 single-bit errors to convert one into the other. With an m-bit memory word, all 2m bit patterns are legal, but due to the way the check bits are computed, only 2m of the 2n codewords are valid. If a memory read turns up an invalid codeword, the computer knows that a memory error has occurred. Given the algorithm for computing the check bits, it is possible to construct a complete list of the legal codewords, and from this list find the two codewords whose Hamming distance is minimum. This distance is the Hamming distance of the complete code. The error-detecting and error-correcting properties of a code depend on its Hamming distance. To detect d single-bit errors, you need a distance d + 1 code because with such a code there is no way that d single-bit errors can change a valid codeword into another valid codeword. Similarly, to correct d single-bit errors, you need a distance 2d + 1 code because that way the legal codewords are so far apart that even with d changes, the original codeword is still closer than any other codeword, so it can be uniquely determined. As a simple example of an error-detecting code, consider a code in which a single parity bit is appended to the data. The parity bit is chosen so that the number of 1 bits in the codeword is even (or odd). Such a code has a distance 2, since any single-bit error produces a codeword with the wrong parity. In other words, it takes two single-bit errors to go from a valid codeword to another valid codeword. It can be used to detect single errors. Whenever a word containing the wrong parity is read from memory, an error condition is signaled. The program cannot continue, but at least no incorrect results are computed. As a simple example of an error-correcting code, consider a code with only four valid codewords: 0000000000, 0000011111, 1111100000, and 1111111111

This code has a distance 5, which means that it can correct double errors. If the codeword 0000000111 arrives, the receiver knows that the original must have been 0000011111 (if there was no more than a double error). If, however, a triple error changes 0000000000 into 0000000111, the error cannot be corrected. Imagine that we want to design a code with m data bits and r check bits that will allow all single-bit errors to be corrected. Each of the 2m legal memory words has n illegal codewords at a distance 1 from it. These are formed by systematically inverting each of the n bits in the n-bit codeword formed from it. Thus each of the 2m legal memory words requires n + 1 bit patterns dedicated to it (for the n possible errors and correct pattern). Since the total number of bit patterns is 2n we must have (n + 1)2m ≤ 2n . Using n = m + r this requirement becomes (m + r + 1) ≤ 2r . Given m, this puts a lower limit on the number of check bits needed to correct single errors. Figure 2-13 shows the number of check bits required for various memory word sizes.

SEC. 2.2

75

PRIMARY MEMORY # #################################################### ! Word size ! Check bits ! Total size ! Percent overhead ! !# #################################################### ! ! ! ! 8 4 12 50 !# #################################################### ! ! ! ! ! ! ! ! ! 16 5 21 31 !# #################################################### ! ! ! ! !# #################################################### ! ! ! ! 32 6 38 19 ! ! ! ! ! 64 7 71 11 !# #################################################### ! ! ! ! !# #################################################### ! ! ! ! 128 8 136 6 ! ! ! ! ! 256 9 265 4 !# #################################################### ! ! ! ! ! ! ! ! ! 512 10 522 2 !# #################################################### ! ! ! !

Figure 2-13. Number of check bits for a code that can correct a single error.

This theoretical lower limit can be achieved using a method due to Richard Hamming (1950). Before taking a look at Hamming’s algorithm, let us look at a simple graphical representation that clearly illustrates the idea of an errorcorrecting code for 4-bit words. The Venn diagram of Fig. 2-14(a) contains three circles, A, B, and C, which together form seven regions. As an example, let us encode the 4-bit memory word 1100 in the regions AB, ABC, AC, and BC, 1 bit per region (in alphabetical order). This encoding is shown in Fig. 2-14(a). A

A 0

1

1 0

B

C

A

0 1

0 1

1 0

1

0 Parity bits

B (a)

0

C

(b)

1 0

Error

1 1

C

0

B (c)

Figure 2-14. (a) Encoding of 1100. (b) Even parity added. (c) Error in AC.

Next we add a parity bit to each of the three empty regions to produce even parity, as illustrated in Fig. 2-14(b). By definition, the sum of the bits in each of the three circles, A, B, and C, is now an even number. In circle A, we have the four numbers 0, 0, 1, and 1, which add up to 2, an even number. In circle B, the numbers are 1, 1, 0, and 0, which also add up to 2, an even number. Finally, in circle C, we have the same thing. In this example all the circles happen to be the same, but sums of 0 and 4 are also possible in other examples. This figure corresponds to a codeword with 4 data bits and 3 parity bits. Now suppose that the bit in the AC region goes bad, changing from a 0 to a 1, as shown in Fig. 2-14(c). The computer can now see that circles A and C have the wrong (odd) parity. The only single-bit change that corrects them is to restore AC

76

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

back to 0, thus correcting the error. In this way, the computer can repair single-bit memory errors automatically. Now let us see how Hamming’s algorithm can be used to construct errorcorrecting codes for any size memory word. In a Hamming code, r parity bits are added to an m-bit word, forming a new word of length m + r bits. The bits are numbered starting at 1, not 0, with bit 1 the leftmost (high-order) bit. All bits whose bit number is a power of 2 are parity bits; the rest are used for data. For example, with a 16-bit word, 5 parity bits are added. Bits 1, 2, 4, 8, and 16 are parity bits, and all the rest are data bits. In all, the memory word has 21 bits (16 data, 5 parity). We will (arbitrarily) use even parity in this example. Each parity bit checks specific bit positions; the parity bit is set so that the total number of 1s in the checked positions is even. The bit positions checked by the parity bits are Bit 1 checks bits 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21. Bit 2 checks bits 2, 3, 6, 7, 10, 11, 14, 15, 18, 19. Bit 4 checks bits 4, 5, 6, 7, 12, 13, 14, 15, 20, 21. Bit 8 checks bits 8, 9, 10, 11, 12, 13, 14, 15. Bit 16 checks bits 16, 17, 18, 19, 20, 21.

In general, bit b is checked by those bits b 1 , b 2 , ..., b j such that b 1 + b 2 + ... + b j = b. For example, bit 5 is checked by bits 1 and 4 because 1 + 4 = 5. Bit 6 is checked by bits 2 and 4 because 2 + 4 = 6, and so on. Figure 2-15 shows construction of a Hamming code for the 16-bit memory word 1111000010101110. The 21-bit codeword is 001011100000101101110. To see how error correction works, consider what would happen if bit 5 were inverted by an electrical surge on the power line. The new codeword would be 001001100000101101110 instead of 001011100000101101110. The 5 parity bits will be checked, with the following results: Parity bit 1 incorrect (1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21 contain five 1s). Parity bit 2 correct (2, 3, 6, 7, 10, 11, 14, 15, 18, 19 contain six 1s). Parity bit 4 incorrect (4, 5, 6, 7, 12, 13, 14, 15, 20, 21 contain five 1s). Parity bit 8 correct (8, 9, 10, 11, 12, 13, 14, 15 contain two 1s). Parity bit 16 correct (16, 17, 18, 19, 20, 21 contain four 1s).

The total number of 1s in bits 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, and 21 should be an even number because even parity is being used. The incorrect bit must be one of the bits checked by parity bit 1—namely, bit 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, or 21. Parity bit 4 is incorrect, meaning that one of bits 4, 5, 6, 7, 12, 13, 14, 15, 20, or

SEC. 2.2

PRIMARY MEMORY

77

21 is incorrect. The error must be one of the bits in both lists, namely, 5, 7, 13, 15, or 21. However, bit 2 is correct, eliminating 7 and 15. Similarly, bit 8 is correct, eliminating 13. Finally, bit 16 is correct, eliminating 21. The only bit left is bit 5, which is the one in error. Since it was read as a 1, it should be a 0. In this manner, errors can be corrected. Memory word 1111000010101110 0 1

0 2

1 3

0 4

1 5

1 6

1 7

0 8

0 0 0 0 1 0 1 1 0 1 1 1 0 9 10 11 12 13 14 15 16 17 18 19 20 21

Parity bits

Figure 2-15. Construction of the Hamming code for the memory word 1111000010101110 by adding 5 check bits to the 16 data bits.

A simple method for finding the incorrect bit is first to compute all the parity bits. If all are correct, there was no error (or more than one). Then add up all the incorrect parity bits, counting 1 for bit 1, 2 for bit 2, 4 for bit 4, and so on. The resulting sum is the position of the incorrect bit. For example, if parity bits 1 and 4 are incorrect but 2, 8, and 16 are correct, bit 5 (1 + 4) has been inverted.

2.2.5 Cache Memory Historically, CPUs have always been faster than memories. As memories have improved, so have CPUs, preserving the imbalance. In fact, as it becomes possible to put more and more circuits on a chip, CPU designers are using these new facilities for pipelining and superscalar operation, making CPUs go even faster. Memory designers have usually used new technology to increase the capacity of their chips, not the speed, so the problem appears to be getting worse in time. What this imbalance means in practice is that after the CPU issues a memory request, it will not get the word it needs for many CPU cycles. The slower the memory, the more cycles the CPU will have to wait. As we pointed out above, there are two ways to deal with this problem. The simplest way is to just start memory READs when they are encountered but continue executing and stall the CPU if an instruction tries to use the memory word before it has arrived. The slower the memory, the more often this problem will occur and the greater the penalty when it does occur. For example, if the memory delay is 10 cycles, it is very likely that one of the next 10 instructions will try to use the word read. The other solution is to have machines that do not stall but instead require the compilers not to generate code to use words before they have arrived. The trouble

78

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

is that this approach is far easier said than done. Often after a LOAD there is nothing else to do, so the compiler is forced to insert NOP (no operation) instructions, which do nothing but occupy a slot and waste time. In effect, this approach is a software stall instead of a hardware stall, but the performance degradation is the same. Actually, the problem is not technology, but economics. Engineers know how to build memories that are as fast as CPUs, but to run at full speed, they have to be located on the CPU chip (because going over the bus to memory is very slow). Putting a large memory on the CPU chip makes it bigger, which makes it more expensive, and even if cost were not an issue, there are limits to how big a CPU chip can be made. Thus the choice comes down to having a small amount of fast memory or a large amount of slow memory. What we would prefer is a large amount of fast memory at a low price. Interestingly enough, techniques are known for combining a small amount of fast memory with a large amount of slow memory to get the speed of the fast memory (almost) and the capacity of the large memory at a moderate price. The small, fast memory is called a cache (from the French cacher, meaning to hide, and pronounced ‘‘cash’’). Below we will briefly describe how caches are used and how they work. A more detailed description will be in Chap. 4. The basic idea behind a cache is simple: the most heavily used memory words are kept in the cache. When the CPU needs a word, it first looks in the cache. Only if the word is not there does it go to main memory. If a substantial fraction of the words are in the cache, the average access time can be greatly reduced. Success or failure thus depends on what fraction of the words are in the cache. For years, people have known that programs do not access their memories completely at random. If a given memory reference is to address A, it is likely that the next memory reference will be in the general vicinity of A. A simple example is the program itself. Except for branches and procedure calls, instructions are fetched from consecutive locations in memory. Furthermore, most program execution time is spent in loops, in which a limited number of instructions are executed over and over. Similarly, a matrix manipulation program is likely to make many references to the same matrix before moving on to something else. The observation that the memory references made in any short time interval tend to use only a small fraction of the total memory is called the locality principle and forms the basis for all caching systems. The general idea is that when a word is referenced, it and some of its neighbors are brought from the large slow memory into the cache, so that the next time it is used, it can be accessed quickly. A common arrangement of the CPU, cache, and main memory is illustrated in Fig. 2-16. If a word is read or written k times in a short interval, the computer will need 1 reference to slow memory and k − 1 references to fast memory. The larger k is, the better the overall performance. We can formalize this calculation by introducing c, the cache access time, m, the main memory access time, and h, the hit ratio, which is the fraction of all

SEC. 2.2

79

PRIMARY MEMORY Main memory

CPU Cache

Bus

Figure 2-16. The cache is logically between the CPU and main memory. Physically, there are several possible places it could be located.

references that can be satisfied out of the cache. In our little example of the previous paragraph, h = (k − 1)/k. Some authors also define the miss ratio, which is 1 − h. With these definitions, we can calculate the mean access time as follows: mean access time = c + (1 − h) m

As h → 1, all references can be satisfied out of the cache, and the access time approaches c. On the other hand, as h → 0, a memory reference is needed every time, so the access time approaches c + m, first a time c to check the cache (unsuccessfully), and then a time m to do the memory reference. On some systems, the memory reference can be started in parallel with the cache search, so that if a cache miss occurs, the memory cycle has already been started. However, this strategy requires that the memory can be stopped in its tracks on a cache hit, making the implementation more complicated. Using the locality principle as a guide, main memories and caches are divided up into fixed-size blocks. When talking about these blocks inside the cache, they are commonly referred to as cache lines. When a cache miss occurs, the entire cache line is loaded from the main memory into the cache, not just the word needed. For example, with a 64-byte line size, a reference to memory address 260 will pull the line consisting of bytes 256 to 319 into one cache line. With a little bit of luck, some of the other words in the cache line will be needed shortly. Operating this way is more efficient than fetching individual words because it is faster to fetch k words all at once than one word k times. Also, having cache entries be more than one word means there are fewer of them, hence a smaller overhead is required. Cache design is an increasingly important subject for high-performance CPUs. One issue is cache size. The bigger the cache, the better it performs, but also the more it costs. A second issue is the size of the cache line. A 16-KB cache can be divided up into 1024 lines of 16 bytes, 2048 lines of 8 bytes, and other combinations. A third issue is how the cache is organized, that is, how does the cache keep track of which memory words are currently being held? We will examine caches in detail in Chap. 4.

80

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

A fourth design issue is whether instructions and data are kept in the same cache or different ones. Having a unified cache (instructions and data use the same cache) is a simpler design and automatically balances instruction fetches against data fetches. Nevertheless, the trend these days is toward a split cache, with instructions in one cache and data in the other. This design is also called a Harvard architecture, the reference going all the way back to Howard Aiken’s Mark III computer, which had different memories for instructions and data. The force driving designers in this direction is the widespread use of pipelined CPUs. The instruction fetch unit needs to access instructions at the same time the operand fetch unit needs access to data. A split cache allows parallel accesses; a unified one does not. Also, since instructions are not modified during execution, the contents of the instruction cache never has to be written back into memory. Finally, a fifth issue is the number of caches. It is not uncommon these days to have chips with a primary cache on chip, a secondary cache off chip but in the same package as the CPU chip, and a third cache still further away.

2.2.6 Memory Packaging and Types From the early days of semiconductor memory until the early 1990s, memory was manufactured, bought, and installed as single chips. Chip densities went from 1K bits to 1M bits and beyond, but each chip was sold as a separate unit. Early PCs often had empty sockets into which additional memory chips could be plugged, if and when the purchaser needed them. At present, a different arrangement is used. A group of chips, typically 8 or 16, is mounted on a tiny printed circuit board and sold as a unit. This unit is called a SIMM (Single Inline Memory Module) or a DIMM (Dual Inline Memory Module), depending on whether it has a row of connectors on one side or both sides of the board. SIMMs have one edge connector with 72 contacts and transfer 32 bits per clock cycle. DIMMs usually have edge connectors with 84 contacts on each side of the board, for a total of 168 contacts and transfer 64 bits per clock cycle. An example SIMM is illustrated in Fig. 2-17. 32-MB memory chip Connector

Figure 2-17. A single inline memory module (SIMM) holding 256 MB. Two of the chips control the SIMM.

A typical SIMM or DIMM configuration might have eight data chips with 256 megabits (32 MB) each. The entire module would then hold 256 MB. Many

SEC. 2.2

PRIMARY MEMORY

81

computers have room for four modules, giving a total capacity of 1 GB when using 256-MB modules and more when using larger ones. A physically smaller DIMM, called an SO-DIMM (Small Outline DIMM) is used in notebook computers. SIMMs and DIMMS can have a parity bit or error correction added, but since the average error rate of a module is one error every 10 years, for most garden-variety computers, error detection and correction are omitted.

2.3 SECONDARY MEMORY No matter how big the main memory is, it is always way too small. People always want to store more information than it can hold, primarily because as technology improves, people begin thinking about storing things that were previously entirely in the realm of science fiction. For example, as the U.S. government’s budget discipline forces government agencies to generate their own revenue, one can imagine the Library of Congress deciding to digitize and sell its full contents as a consumer article (‘‘All of human knowledge for only $99.95’’). Roughly 50 million books, each with 1 MB of text and 1 MB of compressed pictures, requires storing 1014 bytes or 100 terabytes. Storing all 50,000 movies ever made is also in this general ballpark. This amount of information is not going to fit in main memory, at least not for a few decades.

2.3.1 Memory Hierarchies The traditional solution to storing a great deal of data is a memory hierarchy, as illustrated in Fig. 2-18. At the top are the CPU registers, which can be accessed at full CPU speed. Next comes the cache memory, which is currently on the order of 32 KB to a few megabytes. Main memory is next, with sizes currently ranging from 16 MB for entry-level systems to tens of gigabytes at the high end. After that come magnetic disks, the current work horse for permanent storage. Finally, we have magnetic tape and optical disks for archival storage. As we move down the hierarchy, three key parameters increase. First, the access time gets bigger. CPU registers can be accessed in a few nanoseconds. Cache memories take a small multiple of CPU registers. Main memory accesses are typically a few tens of nanoseconds. Now comes a big gap, as disk access times are at least 10 msec, and tape or optical disk access can be measured in seconds if the media have to be fetched and inserted into a drive. Second, the storage capacity increases as we go downward. CPU registers are good for perhaps 128 bytes, caches for a few megabytes, main memories for tens to thousands of megabytes, magnetic disks for a few gigabytes to tens of gigabytes. Tapes and optical disks are usually kept off-line, so their capacity is limited only by the owner’s budget.

82

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Registers Cache

Main memory

Magnetic disk

Tape

Optical disk

Figure 2-18. A five-level memory hierarchy.

Third, the number of bits you get per dollar spent increases down the hierarchy. Although the actual prices change rapidly, main memory is measured in dollars/megabyte, magnetic disk storage in pennies/megabyte, and magnetic tape in dollars/gigabyte or less. We have already looked at registers, cache, and main memory. In the following sections we will look at magnetic disks; after that, we will study optical ones. We will not study tapes because they are rarely used except for backup, and there is not a lot to say about them anyway.

2.3.2 Magnetic Disks A magnetic disk consists of one or more aluminum platters with a magnetizable coating. Originally these platters were as much as 50 cm in diameter, but at present they are typically 3 to 12 cm, with disks for notebook computers already under 3 cm and still shrinking. A disk head containing an induction coil floats just over the surface, resting on a cushion of air (except for floppy disks, where it touches the surface). When a positive or negative current passes through the head, it magnetizes the surface just beneath the head, aligning the magnetic particles facing left or facing right, depending on the polarity of the drive current. When the head passes over a magnetized area, a positive or negative current is induced in the head, making it possible to read back the previously stored bits. Thus as the platter rotates under the head, a stream of bits can be written and later read back. The geometry of a disk track is shown in Fig. 2-19. The circular sequence of bits written as the disk makes a complete rotation is called a track. Each of the tracks is divided up into some number of fixed-length

SEC. 2.3

83

SECONDARY MEMORY Intersector gap or ect 1s

ta b 6 da 409

its

Preamb le

Read/write head

ble m ea Pr

Track width is 1–2 microns

E C C

Dire c

tion

of

dis kr ota

40 96 da ta

bit s

ti o n C

Width of 1 bit is 0.1 to 0.2 microns

C

E

Disk arm

Figure 2-19. A portion of a disk track. Two sectors are illustrated.

sectors, typically containing 512 data bytes, preceded by a preamble that allows the head to be synchronized before reading or writing. Following the data is an Error-Correcting Code (ECC), either a Hamming code, or more commonly, a code that can correct multiple errors called a Reed-Solomon code. Between consecutive sectors is a small intersector gap. Some manufacturers quote their disks’ capacities in unformatted state (as if each track contained only data), but a more honest measurement is the formatted capacity, which does not count the preambles, ECCs and gaps as data. The formatted capacity is typically about 15 percent lower than the unformatted capacity. All disks have movable arms that are capable of moving in and out to different radial distances from the spindle about which the platter rotates. At each radial distance, a different track can be written. The tracks are thus a series of concentric circles about the spindle. The width of a track depends on how large the head is and how accurately the head can be positioned radially. With current technology, disks have between 5000 and 10,000 tracks per centimeter, giving track widths in the 1- to 2-micron range (1 micron = 1/1000 mm). It should be noted that a track is not a physical groove in the surface, but simply an annulus (ring) of magnetized material, with small guard areas separating it from the tracks inside and outside it. The linear bit density around the circumference of the track is different from the radial one. It is determined largely by the purity of the surface and air quality. Current disks achieve densities of 50,000 to 100,000 bits/cm. Thus a bit is about 50 times as big in the radial direction as along the circumference. To go to even higher densities, disk manufacturers are developing technologies in which the ‘‘long’’ dimension of the bits are not along the circumference of the disk, but vertically, down into the iron oxide. This technique is called perpendicular recording and will be commercialized soon. In order to achieve high surface and air quality, most disks are sealed at the factory to prevent dust from getting in. Such drives are called Winchester disks.

SEC. 2.3

85

SECONDARY MEMORY

matters is the average sustained rate over a period of seconds, which has to take into account the necessary seeks and rotational delays as well. A little thought and the use of that old high-school math formula for the circumference of a circle, c = 2πr, will reveal that the outer tracks have more linear distance around them than the inner ones do. Since all magnetic disks rotate at a constant angular velocity, no matter where the heads are, this observation creates a problem. In older drives, manufacturers used the maximum possible linear density on the innermost track, and successively lower linear bit densities on tracks further out. If a disk had 18 sectors per track, for example, each one occupied 20 degrees of arc, no matter which cylinder it was in. Nowadays, a different strategy is used. Cylinders are divided into zones (typically 10 to 30 per drive), and the number of sectors per track is increased in each zone moving outward from the innermost track. This change makes keeping track of information harder but increases the drive capacity, which is viewed as more important. All sectors are the same size. A disk with five zones is shown in Fig. 2-21. Sector

Figure 2-21. A disk with five zones. Each zone has many tracks.

Associated with each drive is a disk controller, a chip that controls the drive. Some controllers contain a full CPU. The controller’s tasks include accepting commands from the software, such as READ, WRITE, and FORMAT (writing all the preambles), controlling the arm motion, detecting and correcting errors, and converting 8-bit bytes read from memory into a serial bit stream and vice versa. Some controllers also handle buffering of multiple sectors, caching sectors read

86

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

for potential future use, and remapping bad sectors. This latter function is caused by the existence of sectors with a bad (permanently magnetized) spot. When the controller discovers a bad sector, it replaces it by one of the spare sectors reserved for this purpose within each cylinder or zone.

2.3.3 Floppy Disks With the advent of the personal computer, a way was needed to distribute software. The solution was found in the diskette or floppy disk, a small, removable medium so called because the early ones were physically flexible. The floppy disk was actually invented by IBM for recording maintenance information about its mainframes for the service staff but was quickly seized on by personal computer manufacturers as a convenient way to distribute software for sale. The general characteristics are the same as the disks we have just described, except that unlike hard disks, where the heads float just above the surface on a cushion of rapidly-moving air, floppy disk heads actually touch the diskettes. As a result, both the media and the heads wear out comparatively quickly. To reduce wear and tear, personal computers retract the heads and stop the rotation when a drive is not reading or writing. Consequently, when the next read or write command is given, there is a delay of about half a second while the motor gets up to speed. Floppy disks had a run of about 20 years, but modern computers are usually shipped without them.

2.3.4 IDE Disks Modern personal computer disks evolved from the one in the IBM PC XT, which was a 10-MB Seagate disk controlled by a Xebec disk controller on a plug-in card. The Seagate disk had 4 heads, 306 cylinders, and 17 sectors/track. The controller was capable of handling two drives. The operating system read from and wrote to a disk by putting parameters in CPU registers and then calling the BIOS (Basic Input Output System), located in the PC’s built-in read-only memory. The BIOS issued the machine instructions to load the disk controller registers that initiated transfers. The technology evolved rapidly from having the controller on a separate board, to having it closely integrated with the drives, starting with IDE (Integrated Drive Electronics) drives in the mid 1980s. However, the BIOS calling conventions were not changed for reasons of backward compatibility. These calling conventions addressed sectors by giving their head, cylinder, and sector numbers, with the heads and cylinders numbered starting at 0 and the sectors starting at 1. This choice was probably due to a mistake on the part of the original BIOS programmer, who wrote his masterpiece in 8088 assembler. With 4 bits for the head, 6 bits for the sector, and 10 bits for the cylinder, the maximum drive could have 16 heads, 63 sectors, and 1024 cylinders, for a total of 1,032,192

SEC. 2.3

SECONDARY MEMORY

87

sectors. Such a maximum drive has a capacity of 504 MB, which probably seemed like infinity at the time but certainly does not today. (Would you fault a new machine today that could not handle drives bigger than a petabyte?) Before too long, drives above 504 MB appeared, but with the wrong geometry (e.g., 4 heads, 32 sectors, 2000 cylinders). There was no way for the operating system to address them due to the long-frozen BIOS calling conventions. As a result, disk controllers began to lie, pretending that the geometry was within the BIOS limits but actually remapping the virtual geometry onto the real geometry. Although this approach worked, it wreaked havoc with operating systems that carefully placed data to minimize seek times. Eventually, IDE drives evolved into EIDE drives (Extended IDE), which also support a second addressing scheme called LBA (Logical Block Addressing), which just numbers the sectors starting at 0 up until a maximum of 228 − 1. This scheme requires the controller to convert LBA addresses to head, sector, and cylinder addresses, but it does get beyond the 504-MB limit. Unfortunately, it created a new bottleneck at 228 × 29 bytes (128 GB). In 1994, when the EIDE standard was adopted, nobody could imagine 128 GB disks. Standards committees, like politicians, have a tendency to push problems forward in time so the next committee has to solve them. EIDE drives and controllers also had other improvements as well. For example, EIDE controllers could have two channels, each with a primary and a secondary drive. This arrangement allowed a maximum of four drives per controller. CD-ROM and DVD drives were also supported, and the transfer rate was increased from 4 MB/sec to 16.67 MB/sec. As disk technology continued to improved, the EIDE standard, continued to evolve, but for some reason the successor to EIDE was called ATA-3 (AT Attachment), a reference to the IBM PC/AT (where AT referred to the thenAdvanced Technology of a 16-bit CPU running at 8 MHz). In the next edition, the standard was called ATAPI-4 (ATA Packet Interface) and the speed was increased to 33 MB/sec. In ATAPI-5 it went to 66 MB/sec. By this time, the 128-GB limit imposed by the 28-bit LBA addresses was looming larger and larger, so ATAPI-6 changed the LBA size to 48 bits. The new standard will run into trouble when disks reach 248 × 29 bytes (128 PB). With a 50% annual increase in capacity, the 48-bit limit will probably last until about 2035. To find out how the problem was solved, please consult the 11th edition of this book. The smart money is betting on increasing the LBA size to 64 bits. The ATAPI-6 standard also increased the transfer rate to 100 MB/sec and addressed the issue of disk noise for the first time. The ATAPI-7 standard is a radical break with the past. Instead of increasing the size of the drive connector (to increase the data rate), this standard uses what is called serial ATA to transfer 1 bit at a time over a 7-pin connector at speeds starting at 150 MB/sec and expected to rise over time to 1.5 GB/sec. Replacing the current 80-wire flat cable with a round cable only a few mm thick improves

88

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

airflow within the computer. Also, serial ATA uses 0.5 volts for signaling (compared to 5 volts on ATAPI-6 drives), which reduces power consumption. It is likely that within a few years, all computers will use serial ATA. The issue of power consumption by disks is an increasingly important one, both at the high end, where data centers have vast disk farms as at the low end, where notebooks are power limited (Gurumurthi et al., 2003).

2.3.5 SCSI Disks SCSI disks are not different from IDE disks in terms of how their cylinders, tracks, and sectors are organized, but they have a different interface and much higher transfer rates. SCSI traces its history back to Howard Shugart, the inventor of the floppy disk, whose company introduced the SASI (Shugart Associates System Interface) disk in 1979. After some modification and much discussion, ANSI standardized it in 1986 and changed the name to SCSI (Small Computer System Interface). SCSI is pronounced ‘‘scuzzy.’’ Since then, increasingly faster versions have been standardized under the names Fast SCSI (10 Mhz), Ultra SCSI (20 MHz), Ultra2 SCSI (40 MHz), Ultra3 SCSI (80 MHz), and Ultra4 SCSI (160 MHz). Each of these has a wide (16-bit) version as well. The main combinations are shown in Fig. 2-22. # ############################################## ! ! Data bits ! Bus MHz ! MB/sec ! Name !# ############################################## ! ! ! ! 8 5 5 ! !#SCSI-1 ! ! ! ############################################## ! Fast SCSI ! ! ! 8 10 10 ! !# ############################################## ! ! ! ! !#Wide ! ! Fast SCSI ! 16 10 20 ! ############################################## ! ! ! ! ! SCSI 8 20 20 ! ############################################## !#Ultra ! ! ! !#Wide ! ! Ultra SCSI ! 16 20 40 ! ############################################## ! ! ! ! ! SCSI 8 40 40 ! !#Ultra2 ! ! ! ############################################## ! Wide Ultra2 SCSI ! ! ! 16 40 80 ! !# ############################################## ! ! ! ! SCSI 8 80 80 ! !#Ultra3 ! ! ! ############################################## ! ! ! 160 ! Ultra3 SCSI ! 16 80 #! Wide ############################################## ! ! ! ! ! ! ! 160 ! SCSI 8 160 #! Ultra4 ############################################## ! ! ! ! ! Ultra4 SCSI ! 16 160 !#Wide ############################################## ! ! 320 ! Figure 2-22. Some of the possible SCSI parameters.

Because SCSI disks have high transfer rates, they are the standard disk in most UNIX workstations from Sun, HP, SGI, and other vendors. They are also the standard disk in Macintoshes and high-end Intel PCs, especially network servers. SCSI is more than just a hard disk interface. It is a bus to which a SCSI controller and up to seven devices can be attached. These can include one or more SCSI hard disks, CD-ROMs, CD recorders, scanners, tape units, and other SCSI

84

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

The first such drives (created by IBM) had 30 MB of sealed, fixed storage and 30 MB of removable storage. Supposedly, these 30-30 disks reminded people of the Winchester 30-30 rifles that played a great role in opening the American frontier, and the name ‘‘Winchester’’ stuck. Most disks consist of multiple platters stacked vertically, as depicted in Fig. 2-20. Each surface has its own arm and head. All the arms are ganged together so they move to different radial positions all at once. The set of tracks at a given radial position is called a cylinder. Current PC disks typically have 6 to 12 platters per drive, giving 12 to 24 recording surfaces. Read/write head (1 per surface) Surface 7 Surface 6 Surface 5 Surface 4 Surface 3

Direction of arm motion

Surface 2 Surface 1 Surface 0

Figure 2-20. A disk with four platters.

Disk performance depends on a variety of factors. To read or write a sector, first the arm must be moved to the right radial position. This action is called a seek. Average seek times (between random tracks) range in the 5- to 10-msec range, although seeks between consecutive tracks are now down below 1 msec. Once the head is positioned radially, there is a delay, called the rotational latency, until the desired sector rotates under the head. Most disks rotate at 5400 RPM, 7200 RPM, or 10,800 RPM, so the average delay (half a rotation) is 3 to 6 msec. Transfer time depends on the linear density and rotation speed. With typical transfer rates of 20 to 40 MB/sec, a 512-byte sector takes between 13 and 26 µsec. Consequently, the seek time and rotational latency dominate the transfer time. Reading random sectors all over the disk is clearly an inefficient way to operate. It is worth mentioning that on account of the preambles, the ECCs, the intersector gaps, the seek times, and the rotational latencies, there is a big difference between a drive’s maximum burst rate and its maximum sustained rate. The maximum burst rate is the data rate once the head is over the first data bit. The computer must be able to handle data coming in this fast. However, the drive can only keep up that rate for one sector. For some applications, such as multimedia, what

SEC. 2.3

SECONDARY MEMORY

89

peripherals. Each SCSI device has a unique ID, from 0 to 7 (15 for wide SCSI). Each device has two connectors: one for input and one for output. Cables connect the output of one device to the input of the next one, in series, like a string of cheap Christmas tree lamps. The last device in the string must be terminated to prevent reflections from the ends of the SCSI bus from interfering with other data on the bus. Typically, the controller is on a plug-in card and the start of the cable chain, although this configuration is not strictly required by the standard. The most common cable for 8-bit SCSI has 50 wires, 25 of which are grounds paired one-to-one with the other 25 wires to provide the excellent noise immunity needed for high-speed operation. Of the 25 wires, 8 are for data, 1 is for parity, 9 are for control, and the remainder are for power or are reserved for future use. The 16-bit (and 32-bit) devices need a second cable for the additional signals. The cables may be several meters long, allowing for external drives, scanners, etc. SCSI controllers and peripherals can operate either as initiators or as targets. Usually, the controller, acting as initiator, issues commands to disks and other peripherals acting as targets. These commands are blocks of up to 16 bytes telling the target what to do. Commands and responses occur in phases, using various control signals to delineate the phases and arbitrate bus access when multiple devices are trying to use the bus at the same time. This arbitration is important because SCSI allows all the devices to run at once, potentially greatly improving performance in an environment with multiple processes active at once (e.g., UNIX or Windows XP). IDE and EIDE allow only one active device at a time.

2.3.6 RAID CPU performance has been increasing exponentially over the past decade, roughly doubling every 18 months. Not so with disk performance. In the 1970s, average seek times on minicomputer disks were 50 to 100 msec. Now seek times are 10 msec. In most technical industries (say, automobiles or aviation), a factor of 5 to 10 performance improvement in two decades would be major news, but in the computer industry it is an embarrassment. Thus the gap between CPU performance and disk performance has become much larger over time. As we have seen, parallel processing is often used to speed up CPU performance. It has occurred to various people over the years that parallel I/O might be a good idea too. In their 1988 paper, Patterson et al. suggested six specific disk organizations that could be used to improve disk performance, reliability, or both (Patterson et al., 1988). These ideas were quickly adopted by industry and have led to a new class of I/O device called a RAID. Patterson et al. defined RAID as Redundant Array of Inexpensive Disks, but industry redefined the I to be ‘‘Independent’’ rather than ‘‘Inexpensive’’ (maybe so they could use expensive disks?). Since a villain was also needed (as in RISC versus CISC, also due to Patterson), the bad guy here was the SLED (Single Large Expensive Disk).

90

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

The idea behind a RAID is to install a box full of disks next to the computer, typically a large server, replace the disk controller card with a RAID controller, copy the data over to the RAID, and then continue normal operation. In other words, a RAID should look like a SLED to the operating system but have better performance and better reliability. Since SCSI disks have good performance, low price, and the ability to have up to 7 drives on a single controller (15 for wide SCSI), it is natural that most RAIDs consist of a RAID SCSI controller plus a box of SCSI disks that appear to the operating system as a single large disk. In this way, no software changes are required to use the RAID, a big selling point for many system administrators. In addition to appearing like a single disk to the software, all RAIDs have the property that the data are distributed over the drives, to allow parallel operation. Several different schemes for doing this were defined by Patterson et al., and they are now known as RAID level 0 through RAID level 5. In addition, there are a few other minor levels that we will not discuss. The term ‘‘level’’ is something of a misnomer since there is no hierarchy involved; there are simply six different organizations possible. RAID level 0 is illustrated in Fig. 2-23(a). It consists of viewing the virtual disk simulated by the RAID as being divided up into strips of k sectors each, with sectors 0 to k − 1 being strip 0, sectors k to 2k − 1 as strip 1, and so on. For k = 1, each strip is a sector; for k = 2 a strip is two sectors, etc. The RAID level 0 organization writes consecutive strips over the drives in round robin fashion, as depicted in Fig. 2-23(a) for a RAID with four disk drives. Distributing data over multiple drives like this is called striping. For example, if the software issues a command to read a data block consisting of four consecutive strips starting at a strip boundary, the RAID controller will break this command up into four separate commands, one for each of the four disks, and have them operate in parallel. Thus we have parallel I/O without the software knowing about it. RAID level 0 works best with large requests, the bigger the better. If a request is larger than the number of drives times the strip size, some drives will get multiple requests, so that when they finish the first request they start the second one. It is up to the controller to split the request up and feed the proper commands to the proper disks in the right sequence and then assemble the results in memory correctly. Performance is excellent and the implementation is straightforward. RAID level 0 works worst with operating systems that habitually ask for data one sector at a time. The results will be correct, but there is no parallelism and hence no performance gain. Another disadvantage of this organization is that the reliability is potentially worse than having a SLED. If a RAID consists of four disks, each with a mean time to failure of 20,000 hours, about once every 5000 hours a drive will fail and all the data will be completely lost. A SLED with a mean time to failure of 20,000 hours would be four times more reliable. Because no redundancy is present in this design, it is not really a true RAID.

SEC. 2.3

(a)

(b)

91

SECONDARY MEMORY

Strip 0

Strip 1

Strip 2

Strip 3

Strip 4

Strip 5

Strip 6

Strip 7

Strip 8

Strip 9

Strip 10

Strip 11

Strip 0

Strip 1

Strip 2

Strip 3

Strip 0

Strip 1

Strip 2

Strip 3

Strip 4

Strip 5

Strip 6

Strip 7

Strip 4

Strip 5

Strip 6

Strip 7

Strip 8

Strip 9

Strip 10

Strip 11

Strip 8

Strip 9

Strip 10

Strip 11

Bit 1

Bit 2

Bit 3

Bit 4

Bit 5

Bit 6

Bit 7

RAID level 0

RAID level 2

(c)

Bit 1

Bit 2

Bit 3

Bit 4

Parity RAID level 3

(d)

(e)

(f)

RAID level 1

Strip 0

Strip 1

Strip 2

Strip 3

P0-3

Strip 4

Strip 5

Strip 6

Strip 7

P4-7

Strip 8

Strip 9

Strip 10

Strip 11

P8-11

Strip 0

Strip 1

Strip 2

Strip 3

P0-3

Strip 4

Strip 5

Strip 6

P4-7

Strip 7

Strip 8

Strip 9

P8-11

Strip 10

Strip 11 RAID level 5

Strip 12

P12-15

Strip 13

Strip 14

Strip 15

P16-19

Strip 16

Strip 17

Strip 18

Strip 19

RAID level 4

Figure 2-23. RAID levels 0 through 5. Backup and parity drives are shown shaded.

92

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

The next option, RAID level 1, shown in Fig. 2-23(b), is a true RAID. It duplicates all the disks, so there are four primary disks and four backup disks. On a write, every strip is written twice. On a read, either copy can be used, distributing the load over more drives. Consequently, write performance is no better than for a single drive, but read performance can be up to twice as good. Fault tolerance is excellent: if a drive crashes, the copy is simply used instead. Recovery consists of simply installing a new drive and copying the entire backup drive to it. Unlike levels 0 and 1, which work with strips of sectors, RAID level 2 works on a word basis, possibly even a byte basis. Imagine splitting each byte of the single virtual disk into a pair of 4-bit nibbles, then adding a Hamming code to each one to form a 7-bit word, of which bits 1, 2, and 4 were parity bits. Further imagine that the seven drives of Fig. 2-23(c) were synchronized in terms of arm position and rotational position. Then it would be possible to write the 7-bit Hamming coded word over the seven drives, one bit per drive. The Thinking Machines CM-2 computer used this scheme, taking 32-bit data words and adding 6 parity bits to form a 38-bit Hamming word, plus an extra bit for word parity, and spread each word over 39 disk drives. The total throughput was immense, because in one sector time it could write 32 sectors worth of data. Also, losing one drive did not cause problems, because loss of a drive amounted to losing 1 bit in each 39-bit word read, something the Hamming code could handle on the fly. On the down side, this scheme requires all the drives to be rotationally synchronized, and it only makes sense with a substantial number of drives (even with 32 data drives and 6 parity drives, the overhead is 19 percent). It also asks a lot of the controller, since it must do a Hamming checksum every bit time. RAID level 3 is a simplified version of RAID level 2. It is illustrated in Fig. 2-23(d). Here a single parity bit is computed for each data word and written to a parity drive. As in RAID level 2, the drives must be exactly synchronized, since individual data words are spread over multiple drives. At first thought, it might appear that a single parity bit gives only error detection, not error correction. For the case of random undetected errors, this observation is true. However, for the case of a drive crashing, it provides full 1-bit error correction since the position of the bad bit is known. If a drive crashes, the controller just pretends that all its bits are 0s. If a word has a parity error, the bit from the dead drive must have been a 1, so it is corrected. Although both RAID levels 2 and 3 offer very high data rates, the number of separate I/O requests per second they can handle is no better than for a single drive. RAID levels 4 and 5 work with strips again, not individual words with parity, and do not require synchronized drives. RAID level 4 [see Fig. 2-23(e)] is like RAID level 0, with a strip-for-strip parity written onto an extra drive. For example, if each strip is k bytes long, all the strips are EXCLUSIVE ORed together, resulting in a parity strip k bytes long. If a drive crashes, the lost bytes can be recomputed from the parity drive.

SEC. 2.3

SECONDARY MEMORY

93

This design protects against the loss of a drive but performs poorly for small updates. If one sector is changed, it is necessary to read all the drives in order to recalculate the parity, which then must be rewritten. Alternatively, it can read the old user data and the old parity data and recompute the new parity from them. Even with this optimization, a small update requires two reads and two writes, clearly a bad arrangement. As a consequence of the heavy load on the parity drive, it may become a bottleneck. This bottleneck is eliminated in RAID level 5 by distributing the parity bits uniformly over all the drives, round robin fashion, as shown in Fig. 223(f). However, in the event of a drive crash, reconstructing the contents of the failed drive is a complex process.

2.3.7 CD-ROMs Optical disks were originally developed for recording television programs, but they can be put to more esthetic use as computer storage devices. Due to their large capacity and low price optical disks are widely used for distributing software, books, movies, and data of all kinds, as well as making backups of hard disks. First-generation optical disks were invented by the Dutch electronics conglomerate Philips for holding movies. They were 30 cm across and marketed under the name LaserVision, but they did not catch on, except in Japan. In 1980, Philips, together with Sony, developed the CD (Compact Disc), which rapidly replaced the 33 1/3 RPM vinyl record for music. The precise technical details for the CD were published in an official International Standard (IS 10149), popularly called the Red Book, due to the color of its cover. (International Standards are issued by the International Organization for Standardization, which is the international counterpart of national standards groups like ANSI, DIN, etc. Each one has an IS number.) The point of publishing the disk and drive specifications as an International Standard is to allow CDs from different music publishers and players from different electronics manufacturers to work together. All CDs are 120 mm across and 1.2 mm thick, with a 15-mm hole in the middle. The audio CD was the first successful mass market digital storage medium. They are supposed to last 100 years. Please check back in 2080 for an update on how well the first batch did. A CD is prepared by using a high-power infrared laser to burn 0.8-micron diameter holes in a coated glass master disk. From this master, a mold is made, with bumps where the laser holes were. Into this mold, molten polycarbonate is injected to form a CD with the same pattern of holes as the glass master. Then a thin layer of reflective aluminum is deposited on the polycarbonate, topped by a protective lacquer and finally a label. The depressions in the polycarbonate substrate are called pits; the unburned areas between the pits are called lands.

94

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

When played back, a low-power laser diode shines infrared light with a wavelength of 0.78 micron on the pits and lands as they stream by. The laser is on the polycarbonate side, so the pits stick out in the direction of the laser as bumps in the otherwise flat surface. Because the pits have a height of one-quarter the wavelength of the laser light, light reflecting off a pit is half a wavelength out of phase with light reflecting off the surrounding surface. As a result, the two parts interfere destructively and return less light to the player’s photodetector than light bouncing off a land. This is how the player tells a pit from a land. Although it might seem simplest to use a pit to record a 0 and a land to record a 1, it is more reliable to use a pit/land or land/pit transition for a 1 and its absence as a 0, so this scheme is used. The pits and lands are written in a single continuous spiral starting near the hole and working out a distance of 32 mm toward the edge. The spiral makes 22,188 revolutions around the disk (about 600 per mm). If unwound, it would be 5.6 km long. The spiral is illustrated in Fig. 2-24. Spiral groove

Pit Land

2K block of user data

Figure 2-24. Recording structure of a Compact Disc or CD-ROM.

To make the music play at a uniform rate, it is necessary for the pits and lands to stream by at a constant linear velocity. Consequently, the rotation rate of the CD must be continuously reduced as the reading head moves from the inside of the CD to the outside. At the inside, the rotation rate is 530 RPM to achieve the desired streaming rate of 120 cm/sec; at the outside it has to drop to 200 RPM to give the same linear velocity at the head. A constant linear velocity drive is quite different than a magnetic disk drive, which operates at a constant angular velocity, independent of where the head is currently positioned. Also, 530 RPM is a far cry from the 3600 to 7200 RPM that most magnetic disks whirl at.

SEC. 2.3

SECONDARY MEMORY

95

In 1984, Philips and Sony realized the potential for using CDs to store computer data, so they published the Yellow Book defining a precise standard for what are now called CD-ROMs (Compact Disc-Read Only Memory). To piggyback on the by-then already substantial audio CD market, CD-ROMs were to be the same physical size as audio CDs, mechanically and optically compatible with them, and produced using the same polycarbonate injection molding machines. The consequences of this decision were that slow variable-speed motors were required, but also that the manufacturing cost of a CD-ROM would be well under one dollar in moderate volume. What the Yellow Book defined was the formatting of the computer data. It also improved the error-correcting abilities of the system, an essential step because although music lovers do not mind losing a bit here and there, computer lovers tend to be Very Picky about that. The basic format of a CD-ROM consists of encoding every byte in a 14-bit symbol. As we saw above, 14 bits is enough to Hamming encode an 8-bit byte with 2 bits left over. In fact, a more powerful encoding system is used. The 14-to-8 mapping for reading is done in hardware by table lookup. At the next level up, a group of 42 consecutive symbols forms a 588-bit frame. Each frame holds 192 data bits (24 bytes). The remaining 396 bits are used for error correction and control. So far, this scheme is identical for audio CDs and CD-ROMs. What the Yellow Book adds is the grouping of 98 frames into a CD-ROM sector, as shown in Fig. 2-25. Every CD-ROM sector begins with a 16-byte preamble, the first 12 of which are 00FFFFFFFFFFFFFFFFFFFF00 (hexadecimal), to allow the player to recognize the start of a CD-ROM sector. The next 3 bytes contain the sector number, needed because seeking on a CD-ROM with its single data spiral is much more difficult than on a magnetic disk with its uniform concentric tracks. To seek, the software in the drive calculates approximately where to go, moves the head there, and then starts hunting around for a preamble to see how good its guess was. The last byte of the preamble is the mode. The Yellow Book defines two modes. Mode 1 uses the layout of Fig. 2-25, with a 16-byte preamble, 2048 data bytes, and a 288-byte error-correcting code (a cross-interleaved Reed-Solomon code). Mode 2 combines the data and ECC fields into a 2336-byte data field for those applications that do not need (or cannot afford the time to perform) error correction, such as audio and video. Note that to provide excellent reliability, three separate error-correcting schemes are used: within a symbol, within a frame, and within a CD-ROM sector. Single-bit errors are corrected at the lowest level, short burst errors are corrected at the frame level, and any residual errors are caught at the sector level. The price paid for this reliability is that it takes 98 frames of 588 bits (7203 bytes) to carry a single 2048-byte payload, an efficiency of only 28 percent. Single-speed CD-ROM drives operate at 75 sectors/sec, which gives a data rate of 153,600 bytes/sec in mode 1 and 175,200 bytes/sec in mode 2. Double-

96

COMPUTER SYSTEMS ORGANIZATION

…

CHAP. 2

Symbols of 14 bits each

42 Symbols make 1 frame Frames of 588 bits, each containing 24 data bytes

… Preamble

Bytes 16

98 Frames make 1 sector Data

ECC

2048

288

Mode 1 sector (2352 bytes)

Figure 2-25. Logical data layout on a CD-ROM.

speed drives are twice as fast, and so on up to the highest speed. A standard audio CD has room for 74 minutes of music, which, if used for mode 1 data, gives a capacity of 681,984,000 bytes. This figure is usually reported as 650 MB because 1 MB is 220 bytes (1,048,576 bytes), not 1,000,000 bytes. Note that even a 32x CD-ROM drive (4,915,200 bytes/sec) is no match for a fast SCSI-2 magnetic disk drive at 10 MB/sec, even though many CD-ROM drives use the SCSI interface (IDE CD-ROM drives also exist). When you realize that the seek time is often several hundred milliseconds, it should be clear that CD-ROM drives are not at all in the same performance category as magnetic disk drives, despite their large capacity. In 1986, Philips struck again with the Green Book, adding graphics and the ability to interleave audio, video and data in the same sector, a feature essential for multimedia CD-ROMs. The last piece of the CD-ROM puzzle is the file system. To make it possible to use the same CD-ROM on different computers, agreement was needed on CDROM file systems. To get this agreement, representatives of many computer companies met at Lake Tahoe in the High Sierras on the California-Nevada boundary and devised a file system that they called High Sierra. It later evolved into an International Standard (IS 9660). It has three levels. Level 1 uses file names of up to 8 characters optionally followed by an extension of up to 3 characters (the MS-DOS file naming convention). File names may contain only uppercase letters, digits, and the underscore. Directories may be nested up to eight deep, but directory names may not contain extensions. Level 1 requires all files to be contiguous, which is not a problem on a medium written only once. Any CD-ROM conformant to IS 9660 level 1 can be read using MS-DOS, an Apple computer, a UNIX computer, or just about any other computer. CD-ROM publishers regard this property as being a big plus. IS 9660 level 2 allows names up to 32 characters, and level 3 allows noncontiguous files. The Rock Ridge extensions (whimsically named after the town in

SEC. 2.3

SECONDARY MEMORY

97

the Mel Brooks film Blazing Saddles) allow very long names (for UNIX), UIDs, GIDs, and symbolic links, but CD-ROMs not conforming to level 1 will not be readable on all computers.

2.3.8 CD-Recordables Initially, the equipment needed to produce a master CD-ROM (or audio CD, for that matter) was extremely expensive. But as usual in the computer industry, nothing stays expensive for long. By the mid 1990s, CD recorders no bigger than a CD player were a common peripheral available in most computer stores. These devices were still different from magnetic disks because once written, CD-ROMs could not be erased. Nevertheless, they quickly found a niche as a backup medium for large hard disks and also allowed individuals or startup companies to manufacture their own small-run CD-ROMs or make masters for delivery to high-volume commercial CD duplication plants. These drives are known as CDRs (CD-Recordables). Physically, CD-Rs start with 120-mm polycarbonate blanks that are like CDROMs, except that they contain a 0.6-mm wide groove to guide the laser for writing. The groove has a sinusoidal excursion of 0.3 mm at a frequency of exactly 22.05 kHz to provide continuous feedback so the rotation speed can be accurately monitored and adjusted if need be. The first CD-Rs looked like regular CDROMs, except that they were gold colored on top instead of silver colored. The gold color came from the use of real gold instead of aluminum for the reflective layer. Unlike silver CDs, which have physical depressions on them, on CD-Rs the differing reflectivity of pits and lands has to be simulated. This is done by adding a layer of dye between the polycarbonate and the reflective layer, as shown in Fig. 2-26. Two kinds of dye are used: cyanine, which is green, and pthalocyanine, which is a yellowish orange. Chemists can argue endlessly about which one is better. These dyes are similar to those used in photography, which explains why Kodak and Fuji are major manufacturers of CD-Rs. Eventually, an aluminum reflective layer replaced the gold one. In its initial state, the dye layer is transparent and lets the laser light pass through and reflect off the reflective layer. To write, the CD-R laser is turned up to high power (8–16 mW). When the beam hits a spot of dye, it heats up, breaking a chemical bond. This change to the molecular structure creates a dark spot. When read back (at 0.5 mW), the photodetector sees a difference between the dark spots where the dye has been hit and transparent areas where it is intact. This difference is interpreted as the difference between pits and lands, even when read back on a regular CD-ROM reader or even on an audio CD player. No new kind of CD could hold up its head with pride without a colored book, so CD-R has the Orange Book, published in 1989. This document defines CD-R and also a new format, CD-ROM XA, which allows CD-Rs to be written incrementally, a few sectors today, a few tomorrow, and a few next month. A group of consecutive sectors written at once is called a CD-ROM track.

98

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Printed label Protective lacquer Reflective layer layer Dye

Dark spot in the dye layer burned by laser when writing

1.2 mm Polycarbonate Direction of motion Photodetector

Substrate Lens Prism Infrared laser diode

Figure 2-26. Cross section of a CD-R disk and laser (not to scale). A CD-ROM has a similar structure, except without the dye layer and with a pitted aluminum layer instead of a reflective layer.

One of the first uses of CD-R was for the Kodak PhotoCD. In this system the customer brings a roll of exposed film and his old PhotoCD to the photo processor and gets back the same PhotoCD with the new pictures added after the old ones. The new batch, which is created by scanning in the negatives, is written onto the PhotoCD as a separate CD-ROM track. Incremental writing is needed because the CD-R blanks are too expensive to provide a new one for every film roll. However, incremental writing creates a new problem. Prior to the Orange Book, all CD-ROMs had a single VTOC (Volume Table of Contents) at the start. That scheme does not work with incremental (i.e., multitrack) writes. The Orange Book’s solution is to give each CD-ROM track its own VTOC. The files listed in the VTOC can include some or all of the files from previous tracks. After the CD-R is inserted into the drive, the operating system searches through all the CD-ROM tracks to locate the most recent VTOC, which gives the current status of the disk. By including some, but not all, of the files from previous tracks in the current VTOC, it is possible to give the illusion that files have been deleted. Tracks can be grouped into sessions, leading to multisession CD-ROMs. Standard audio CD players cannot handle multisession CDs since they expect a single VTOC at the start. Each track has to be written in a single continuous operation without stopping. As a consequence, the hard disk from which the data are coming has to be fast enough to deliver it on time. If the files to be copied are spread all over the hard disk, the seek times may cause the data stream to the CD-R to dry up and cause a dreaded buffer underrun. A buffer underrun results in producing a nice shiny (but

SEC. 2.3

SECONDARY MEMORY

99

somewhat expensive) coaster for your drinks, or a 120-mm gold- or silver-colored frisbee. CD-R software usually offers the option of collecting all the input files into a single contiguous 650-MB CD-ROM image prior to burning the CD-R, but this process typically doubles the effective writing time, requires 650 MB of free disk space, and still does not protect against hard disks that panic and decide to do a thermal recalibration when they get too hot. CD-R makes it possible for individuals and companies to easily copy CDROMs (and audio CDs), generally in violation of the publisher’s copyright. Several schemes have been devised to make such piracy harder and to make it difficult to read a CD-ROM using anything other than the publisher’s software. One of them involves recording all the file lengths on the CD-ROM as multigigabyte, thwarting any attempts to copy the files to hard disk using standard copying software. The true lengths are embedded in the publisher’s software or hidden (possibly encrypted) on the CD-ROM in an unexpected place. Another scheme uses intentionally incorrect ECCs in selected sectors, in the expectation that CD copying software will ‘‘fix’’ the errors. The application software checks the ECCs itself, refusing to work if they are correct. Using nonstandard gaps between the tracks and other physical ‘‘defects’’ are also possibilities.

2.3.9 CD-Rewritables Although people are used to other write-once media such as paper and photographic film, there is a demand for a rewritable CD-ROM. One technology now available is CD-RW (CD-ReWritable), which uses the same size media as CDR. However, instead of cyanine or pthalocyanine dye, CD-RW uses an alloy of silver, indium, antimony, and tellurium for the recording layer. This alloy has two stable states: crystalline and amorphous, with different reflectivities. CD-RW drives use lasers with three different powers. At high power, the laser melts the alloy, converting it from the high-reflectivity crystalline state to the low-reflectivity amorphous state to represent a pit. At medium power, the alloy melts and reforms in its natural crystalline state to become a land again. At low power, the state of the material is sensed (for reading), but no phase transition occurs. The reason CD-RW has not replaced CD-R completely is that the CD-RW blanks are more expensive than the CD-R blanks. Also, for applications consisting of backing up hard disks, the fact that once written, a CD-R cannot be accidentally erased is a big plus.

2.3.10 DVD The basic CD/CD-ROM format has been around since 1980. The technology has improved since then, so higher-capacity optical disks are now economically feasible and there is great demand for them. Hollywood would dearly love to

100

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

replace analog video tapes by digital disks, since disks have a higher quality, are cheaper to manufacture, last longer, take up less shelf space in video stores, and do not have to be rewound. The consumer electronics companies are looking for a new blockbuster product, and many computer companies want to add multimedia features to their software. This combination of technology and demand by three immensely rich and powerful industries has led to DVD, originally an acronym for Digital Video Disk, but now officially Digital Versatile Disk. DVDs use the same general design as CDs, with 120-mm injection-molded polycarbonate disks containing pits and lands that are illuminated by a laser diode and read by a photodetector. What is new is the use of 1. Smaller pits (0.4 microns versus 0.8 microns for CDs). 2. A tighter spiral (0.74 microns between tracks versus 1.6 microns for CDs). 3. A red laser (at 0.65 microns versus 0.78 microns for CDs). Together, these improvements raise the capacity sevenfold, to 4.7 GB. A 1x DVD drive operates at 1.4 MB/sec (versus 150 KB/sec for CDs). Unfortunately, the switch to the red lasers used in supermarkets means that DVD players will require a second laser or fancy conversion optics to be able to read existing CDs and CDROMs, something not all of them may provide. Also, reading CD-Rs and CDRWs on a DVD drive may not be possible. Is 4.7 GB enough? Maybe. Using MPEG-2 compression (standardized in IS 13346), a 4.7 GB DVD disk can hold 133 minutes of full-screen, full-motion video at high resolution (720 × 480), as well as soundtracks in up to eight languages and subtitles in 32 more. About 92 percent of all the movies Hollywood has ever made are under 133 minutes. Nevertheless, some applications such as multimedia games or reference works may need more, and Hollywood would like to put multiple movies on the same disk, so four formats have been defined: 1. Single-sided, single-layer (4.7 GB). 2. Single-sided, dual-layer (8.5 GB). 3. Double-sided, single-layer (9.4 GB). 4. Double-sided, dual-layer (17 GB). Why so many formats? In a word: politics. Philips and Sony wanted singlesided, dual-layer disks for the high capacity version, but Toshiba and Time Warner wanted double-sided, single-layer disks. Philips and Sony did not think people would be willing to turn the disks over, and Time Warner did not believe putting two layers on one side could be made to work. The compromise: all combinations, but the market will determine which ones survive.

SEC. 2.3

SECONDARY MEMORY

101

The dual layering technology has a reflective layer at the bottom, topped with a semireflective layer. Depending on where the laser is focused, it bounces off one layer or the other. The lower layer needs slightly larger pits and lands to be read reliably, so its capacity is slightly smaller than the upper layer’s. Double-sided disks are made by taking two 0.6-mm single-sided disks and gluing them together back to back. To make the thicknesses of all versions the same, a single-sided disk consists of a 0.6-mm disk bonded to a blank substrate (or perhaps in the future, one consisting of 133 minutes of advertising, in the hope that people will be curious as to what is down there). The structure of the double-sided, dual-layer disk is illustrated in Fig. 2-27. 0.6 mm Single-sided disk

Polycarbonate substrate 1

Semireflective layer Aluminum reflector

Adhesive layer 0.6 mm Single-sided disk

Aluminum reflector Polycarbonate substrate 2

Semireflective layer

Figure 2-27. A double-sided, dual layer DVD disk.

DVD was devised by a consortium of 10 consumer electronics companies, seven of them Japanese, in close cooperation with the major Hollywood studios (some of which are owned by the Japanese electronics companies in the consortium). The computer and telecommunications industries were not invited to the picnic, and the resulting focus was on using DVD for movie rental and sales shows. For example, standard features include real-time skipping of dirty scenes (to allow parents to turn a film rated NC17 into one safe for toddlers), six-channel sound, and support for Pan-and-Scan. The latter feature allows the DVD player to dynamically decide how to crop the left and right edges off movies (whose width:height ratio is 3:2) to fit on current television sets (whose aspect ratio is 4:3). Another item the computer industry probably would not have thought of is an intentional incompatibility between disks intended for the United States and disks intended for Europe and yet other standards for other continents. Hollywood demanded this ‘‘feature’’ because new films are always released first in the United States and then shipped to Europe when the videos come out in the United States. The idea was to make sure European video stores could not buy videos in the U.S. too early, thereby reducing new movies’ European theater sales. If Hollywood had been running the computer industry, we would have had 3.5-inch floppy disks in the United States and 9-cm floppy disks in Europe.

102

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

2.3.11 Blu-Ray Nothing stands still in the computer business, certainly not storage technology. DVD was barely introduced before its successor threatened to make it obsolete. The successor to DVD is Blu-Ray, so called because it uses a blue laser instead of the red one used by DVDs. A blue laser has a shorter wavelength than a red one, which allows it to focus more accurately and thus support smaller pits and lands. Single-sided Blu-Ray disks hold about 25 GB of data; double-sided ones hold about 50 GB. The data rate is about 4.5 MB/sec, which is good for an optical disk, but still insignificant compared to magnetic disks (cf. ATAPI-6 at 100 MB/sec and wide Ultra4 SCSI at 320 MB/sec). It is expected that Blu-Ray will eventually replace CD-ROMs and DVDs, but this transition will take some years.

2.4 INPUT/OUTPUT As we mentioned at the start of this chapter, a computer system has three major components: the CPU, the memories (primary and secondary), and the I/O (Input/Output) equipment such as printers, scanners, and modems. So far we have looked at the CPU and the memories. Now it is time to examine the I/O equipment and how it is connected to the rest of the system.

2.4.1 Buses Physically, most personal computers and workstations have a structure similar to the one shown in Fig. 2-28. The usual arrangement is a metal box with a large printed circuit board at the bottom, called the motherboard (parentboard, for the politically correct). The motherboard contains the CPU chip, some slots into which DIMM modules can be clicked, and various support chips. It also contains a bus etched along its length, and sockets into which the edge connectors of I/O boards can be inserted (the PCI bus). Older PCs also have a second bus (the ISA bus), for legacy I/O boards, but modern computers usually lack it and it is rapidly dying off. The logical structure of a simple low-end personal computer is shown in Fig. 2-29. This one has a single bus used to connect the CPU, memory, and I/O devices; most systems have two or more buses. Each I/O device consists of two parts: one containing most of the electronics, called the controller, and one containing the I/O device itself, such as a disk drive. The controller is usually contained on a board plugged into a free slot, except for those controllers that are not optional (such as the keyboard), which are sometimes located on the motherboard. Even though the display (monitor) is not an option, the video controller is sometimes located on a plug-in board to allow the user to choose between boards with

SEC. 2.4

103

INPUT/OUTPUT SCSI controller Sound card

Modem

Card cage Edge connector

Figure 2-28. Physical structure of a personal computer.

or without graphics accelerators, extra memory, and so on. The controller connects to its device by a cable attached to a connector on the back of the box. Monitor

CPU

Memory

Video controller

Keyboard

CD-ROM drive

Hard disk drive

Keyboard controller

CD-ROM controller

Hard disk controller

Bus

Figure 2-29. Logical structure of a simple personal computer.

The job of a controller is to control its I/O device and handle bus access for it. When a program wants data from the disk for example, it gives a command to the disk controller, which then issues seeks and other commands to the drive. When the proper track and sector have been located, the drive begins outputting the data as a serial bit stream to the controller. It is the job of the controller to break the bit stream up into units, and write each unit into memory, as it is assembled. A unit is typically one or more words. A controller that reads or writes data to or from memory without CPU intervention is said to be performing Direct Memory Access, better known by its acronym DMA. When the transfer is completed, the controller normally causes an interrupt, forcing the CPU to immediately suspend running its current program and start running a special procedure, called an

104

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

interrupt handler, to check for errors, take any special action needed, and inform the operating system that the I/O is now finished. When the interrupt handler is finished, the CPU continues with the program that was suspended when the interrupt occurred. The bus is not only used by the I/O controllers, but also by the CPU for fetching instructions and data. What happens if the CPU and an I/O controller want to use the bus at the same time? The answer is that a chip called a bus arbiter decides who goes next. In general, I/O devices are given preference over the CPU, because disks and other moving devices cannot be stopped, and forcing them to wait would result in lost data. When no I/O is in progress, the CPU can have all the bus cycles for itself to reference memory. However, when some I/O device is also running, that device will request and be granted the bus when it needs it. This process is called cycle stealing and it slows down the computer. This design worked fine for the first personal computers, since all the components were roughly in balance. However, as the CPUs, memories, and I/O devices got faster, a problem arose: the bus could no longer handle the load presented. On a closed system, such as an engineering workstation, the solution was to design a new and faster bus for the next model. Because nobody ever moved I/O devices from an old model to a new one, this approached worked fine. However, in the PC world, people often upgraded their CPU but wanted to move their printer, scanner, and modem to the new system. Also, a huge industry had grown up around providing a vast range of I/O devices for the IBM PC bus, and this industry had exceedingly little interest in throwing out its entire investment and starting over. IBM learned this the hard way when it brought out the successor to the IBM PC, the PS/2 range. The PS/2 had a new, and faster bus, but most clone makers continued to use the old PC bus, now called the ISA (Industry Standard Architecture) bus. Most disk and I/O device makers also continued to make controllers for it, so IBM found itself in the peculiar situation of being the only PC maker that was no longer IBM compatible. Eventually, it was forced back to supporting the ISA bus. As an aside, please note that ISA stands for Instruction Set Architecture in the context of machine levels whereas it stands for Industry Standard Architecture in the context of buses. Nevertheless, despite the market pressure not to change anything, the old bus really was too slow, so something had to be done. This situation led to other companies developing machines with multiple buses, one of which was the old ISA bus, or its backward-compatible successor, the EISA (Extended ISA) bus. The most popular of these now is the PCI (Peripheral Component Interconnect) bus. It was designed by Intel, but Intel decided to put all the patents in the public domain, to encourage the entire industry (including its competitors) to adopt it. The PCI bus can be used in many configurations, but a typical one is illustrated in Fig. 2-30. Here the CPU talks to a memory controller over a dedicated high-speed connection. The controller talks to the memory and to the PCI bus directly, so CPU-memory traffic does not go over the PCI bus. However, high-

SEC. 2.4

105

INPUT/OUTPUT

bandwidth (i.e., high data rate) peripherals can connect to the PCI bus directly. In addition, here the PCI bus has a bridge to the ISA bus, so that ISA controllers and their devices can still be used, although as mentioned earlier, the ISA bus is being phased out. A machine of this design would typically contain three or four empty PCI slots and one or two ISA slots, to allow customers to plug in both old ISA I/O cards (usually for slow devices) and new PCI I/O cards (usually for fast devices).

; ;;

Memory bus

SCSI bus

SCSI scanner

SCSI disk

Sound card

Main memory

PCI bridge

CPU cache

SCSI controller

Printer controller

Video controller

ISA bridge

Network controller PCI bus

Modem

ISA bus

Figure 2-30. A typical modern PC with a PCI bus and an ISA bus. The modem and sound card are ISA devices; the SCSI controller is a PCI device.

Many kinds of I/O devices are available today. A few of the more common ones are discussed below.

2.4.2 Terminals Computer terminals consist of two parts: a keyboard and a monitor. In the mainframe world, these parts are often integrated into a single device and attached to the main computer by a serial line or over a telephone line. In the airline reservation, banking, and other mainframe-oriented industries, these devices are still in widespread use. In the personal computer world, the keyboard and monitor are independent devices. Either way, the technology of the two parts is the same. Keyboards Keyboards come in several varieties. The original IBM PC came with a keyboard that had a snap-action switch under each key that gave tactile feedback and made a click when the key was depressed far enough. Nowadays, the cheaper

106

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

keyboards have keys that just make mechanical contact when depressed. Better ones have a sheet of elastometric material (a kind of rubber) between the keys and the underlying printed circuit board. Under each key is a small dome that buckles when depressed far enough. A small spot of conductive material inside the dome closes the circuit. Some keyboards have a magnet under each key that passes through a coil when struck, thus inducing a current that can be detected. Various other methods, both mechanical and electromagnetic, are also in use. On personal computers, when a key is depressed, an interrupt is generated and the keyboard interrupt handler (a piece of software that is part of the operating system) is started. The interrupt handler reads a hardware register inside the keyboard controller to get the number of the key (1 through 102) that was just depressed. When a key is released, a second interrupt is caused. Thus if a user depresses the SHIFT key, then depresses and releases the M key, then releases the SHIFT key, the operating system can see that the user wants an uppercase ‘‘M’’ rather than a lowercase ‘‘m.’’ Handling of multikey sequences involving SHIFT, CTRL, and ALT is done entirely in software (including the infamous CTRLALT-DEL key sequence that is used to reboot all IBM PCs and clones). CRT Monitors A monitor is a box containing a CRT (Cathode Ray Tube) and its power supplies. The CRT contains a gun that can shoot an electron beam against a phosphorescent screen near the front of the tube, as shown in Fig. 2-31(a). (Color monitors have three electron guns, one each for red, green, and blue.) During the horizontal scan, the beam sweeps across the screen in about 50 µsec, tracing out an almost horizontal line on the screen. Then it executes a horizontal retrace to get back to the left-hand edge in order to begin the next sweep. A device like this that produces an image line by line is called a raster scan device. Horizontal scan Grid Electron gun

Screen Spot on screen Vacuum

Vertical deflection plate

Vertical retrace Horizontal retrace (a)

(b)

Figure 2-31. (a) Cross section of a CRT. (b) CRT scanning pattern.

SEC. 2.4

INPUT/OUTPUT

107

Horizontal sweeping is controlled by a linearly increasing voltage applied to the horizontal deflection plates placed to the left and right of the electron gun. Vertical motion is controlled by a much more slowly linearly increasing voltage applied to the vertical deflection plates placed above and below the gun. After somewhere between 400 and 1000 sweeps, the voltages on the vertical and horizontal deflection plates are rapidly reversed together to put the beam back in the upper left-hand corner. A full-screen image is normally repainted between 30 and 60 times a second. The beam motions are shown in Fig. 2-31(b). Although we have described CRTs as using electric fields for sweeping the beam across the screen, many models use magnetic fields instead of electric ones, especially in high-end monitors. To produce a pattern of dots on the screen, a grid is present inside the CRT. When a positive voltage is applied to the grid, the electrons are accelerated, causing the beam to hit the screen and make it glow briefly. When a negative voltage is used, the electrons are repelled, so they do not pass through the grid and the screen does not glow. Thus the voltage applied to the grid causes the corresponding bit pattern to appear on the screen. This mechanism allows a binary electrical signal to be converted into a visual display consisting of bright and dark spots. Flat Panel Displays CRTs are far too bulky and heavy to be used in notebook computers, so a completely different technology is needed for their screens. The most common one is LCD (Liquid Crystal Display) technology. It is highly complex, has many variations, and is changing rapidly, so this description will, of necessity, be brief and greatly simplified. Liquid crystals are viscous organic molecules that flow like a liquid but also have spatial structure, like a crystal. They were discovered by an Austrian botanist (Rheinitzer) in 1888, and first applied to displays (e.g., calculators, watches) in the 1960s. When all the molecules are lined up in the same direction, the optical properties of the crystal depend on the direction and polarization of the incoming light. Using an applied electric field, the molecular alignment, hence the optical properties, can be changed. In particular, by shining a light through a liquid crystal, the intensity of the light exiting from it can be controlled electrically. This property can be exploited to construct flat panel displays. An LCD display screen consists of two parallel glass plates between which is a sealed volume containing a liquid crystal. Transparent electrodes are attached to both plates. A light behind the rear plate (either natural or artificial) illuminates the screen from behind. The transparent electrodes attached to each plate are used to create electric fields in the liquid crystal. Different parts of the screen get different voltages, to control the image displayed. Glued to the front and rear of the screen are polaroids because the display technology requires the use of polarized light. The general setup is shown in Fig. 2-32(a).

108

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Liquid crystal Rear glass plate Rear electrode

$@;# ;$@#

Rear polaroid

Front glass plate Front electrode Front polaroid

y Dark

z

Bright

Light source

Notebook computer (a)

(b)

Figure 2-32. (a) The construction of an LCD screen. (b) The grooves on the rear and front plates are perpendicular to one another.

Although many kinds of LCD displays are in use, we will now consider one particular kind of display, the TN (Twisted Nematic) display as an example. In this display, the rear plate contains tiny horizontal grooves and the front plate contains tiny vertical grooves, as illustrated in Fig. 2-32(b). In the absence of an electric field, the LCD molecules tend to align with the grooves. Since the front and rear alignments differ by 90 degrees, the molecules (and thus the crystal structure) twist from rear to front. At the rear of the display is a horizontal polaroid. It only allows in horizontally polarized light. At the front of the display is a vertical polaroid. It only allows vertically polarized light to pass through. If there were no liquid present between the plates, horizontally polarized light let in by the rear polaroid would be blocked by the front polaroid, making the screen uniformly black. However the twisted crystal structure of the LCD molecules guides the light as it passes and rotates its polarization, making it come out vertically. Thus in the absence of an electric field, the LCD screen is uniformly bright. By applying a voltage to selected parts of the plate, the twisted structure can be destroyed, blocking the light in those parts. Two schemes can be used for applying the voltage. In a (low-cost) passive matrix display, both electrodes contain parallel wires. In a 640 × 480 display, for example, the rear electrode might have 640 vertical wires and the front one might

SEC. 2.4

INPUT/OUTPUT

109

have 480 horizontal ones. By putting a voltage on one of the vertical wires and then pulsing one of the horizontal ones, the voltage at one selected pixel position can be changed, making it go dark briefly. By repeating this pulse with the next pixel and then the next one, a dark scan line can be painted, analogous to how a CRT works. Normally, the entire screen is painted 60 times a second to fool the eye into thinking there is a constant image there, again, the same way as a CRT. The other scheme in widespread use is the active matrix display. It is more expensive but it gives a better image. Instead of just having two sets of perpendicular wires, it has a tiny switching element at each pixel position on one of the electrodes. By turning these on and off, an arbitrary voltage pattern can be created across the screen, allowing for an arbitrary bit pattern. The switching elements are called thin film transistors and the flat panel displays using them are often called TFT displays. Most notebook computers and stand-alone flat panel displays for desktop computers use TFT technology now. So far we have described how a monochrome display works. Suffice it to say that color displays use the same general principles as monochrome displays, but that the details are a great deal more complicated. Optical filters are used to separate the white light into red, green, and blue components at each pixel position so these can be displayed independently. Every color can be built up from a linear superposition of these three primary colors. Video RAM Both CRTs and TFT displays are refreshed 60–100 times per second from a special memory, called a video RAM, on the display’s controller card. This memory has one or more bit maps that represent the screen. On a screen with, say, 1600 × 1200 picture elements, called pixels, the video RAM would contain 1600 × 1200 values, one for each pixel. In fact, it might contain many such bit maps, to allow rapid switching from one screen image to another. On a high-end display, each pixel would be represented as a 3-byte RGB value, one each for the intensity of the red, green, and blue components of the pixel’s color. From the laws of physics, it is known that any color can be constructed from a linear superposition of red, green, and blue light. A video RAM with 1600 × 1200 pixels at 3 bytes/pixels requires almost 5.5 MB to store the image and a fair amount of CPU time to do anything with it. For this reason, some computers compromise by using an 8-bit number to indicate the color desired. This number is then used as an index into a hardware table, called the color palette that contains 256 entries, each holding a 24-bit RGB value. Such a design, called indexed color, reduces the memory video RAM memory requirements by 2/3, but allows only 256 colors on the screen at once. Usually, each window on the screen has its own mapping, but with only one hardware color palette, often when multiple windows are present on the screen, only the current one has its colors rendered correctly.

110

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Bit-mapped video displays require a lot of bandwidth. To display full-screen, full-color multimedia on a 1600 × 1200 display requires copying 5.5 MB of data to the video RAM for every frame. For full-motion video, a rate of at least 25 frame/sec is needed, for a total data rate of 137.5 MB/sec. This load is far more than the (E)ISA bus can handle, and even more than the original PCI bus could handle (127.2 MB/sec). Of course, smaller images require less bandwidth, but bandwidth is still a major issue. To allow more bandwidth from the CPU to the video RAM, starting with the Pentium II, Intel added support for a new bus to the video RAM, the AGP bus (Accelerated Graphics Port), which can transfer 32 bits at a rate of 66 MHz for a data rate of 252 MB/sec. Subsequent versions ran at 2x, 4x, and even 8x to provide sufficient bandwidth for highly interactive graphics without overloading the main PCI bus.

2.4.3 Mice As time goes on, computers are being used by people with less expertise in how computers work. Computers of the ENIAC generation were used only by the people who built them. In the 1950s, computers were only used by highly-skilled professional programmers. Now, computers are widely used by people who need to get some job done and do not know (or even want to know) much about how computers work or how they are programmed. In the old days, most computers had command line interfaces, to which users typed commands. Since people who are not computer specialists often perceived command line interfaces as user-unfriendly, if not downright hostile, many computer vendors developed point-and-click interfaces, such as the Macintosh and Windows. Using this model requires having a way to point at the screen. The most common way of allowing users to point at the screen is with a mouse. A mouse is a small plastic box that sits on the table next to the keyboard. When it is moved around on the table, a little pointer on the screen moves too, allowing users to point at screen items. The mouse has one, two, or three buttons on top, to allow users to select items from menus. Much blood has been spilled as a result of arguments about how many buttons a mouse ought to have. Naive users prefer one (it is hard to push the wrong button if there is only one), but sophisticated ones like the power of multiple buttons to do fancy things. Three kinds of mice have been produced: mechanical mice, optical mice, and optomechanical mice. The first mice had two rubber wheels protruding through the bottom, with their axles perpendicular to one another. When the mouse was moved parallel to its main axis, one wheel turned. When it is moved perpendicular to its main axis, the other one turned. Each wheel drove a variable resistor or potentiometer. By measuring changes in the resistance, it was possible to see how much each wheel had rotated and thus calculate how far the mouse had moved in

SEC. 2.4

111

INPUT/OUTPUT

each direction. In recent years, this design has largely been replaced by one in which a ball that protrudes slightly from the bottom is used instead of wheels. It is shown in Fig. 2-33. Pointer controlled by mouse Window

Menu

Cut Paste Copy

Mouse buttons Mouse

Rubber ball

Figure 2-33. A mouse being used to point to menu items.

The second kind of mouse is the optical mouse. This kind has no wheels or ball. Instead, it has an LED (Light Emitting Diode) and a photodetector on the bottom. The optical mouse is used on top of a special plastic pad containing a rectangular grid of closely spaced lines. As the mouse moves over the grid, the photodetector senses line crossings by seeing the changes in the amount of light being reflected back from the LED. Electronics inside the mouse count the number of grid lines crossed in each direction. The third kind of mouse is optomechanical. Like the newer mechanical mouse, it has a rolling ball that turns two shafts aligned at 90 degrees to each other. The shafts are connected to encoders that have slits through which light can pass. As the mouse moves, the shafts rotate, and light pulses strike the detectors whenever a slit comes between an LED and its detector. The number of pulses detected is proportional to the amount of motion. Although mice can be set up in various ways, a common arrangement is to have the mouse send a sequence of 3 bytes to the computer every time the mouse moves a certain minimum distance (e.g., 0.01 inch), sometimes called a mickey. Usually, these characters come in on a serial line, one bit at time. The first byte contains a signed integer telling how many units the mouse has moved in the xdirection since the last time. The second byte gives the same information for y motion. The third byte contains the current state of the mouse buttons. Sometimes 2 bytes are used for each coordinate.

112

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

Low-level software in the computer accepts this information as it comes in and converts the relative movements sent by the mouse to an absolute position. It then displays an arrow on the screen at the position corresponding to where the mouse is. When the arrow points at the proper item, the user clicks a mouse button, and the computer can then figure out which item has been selected from its knowledge of where the arrow is on the screen.

2.4.4 Printers Having prepared a document or fetched a page from the World Wide Web, users often want to print it, so all computers can be equipped with a printer. In this section we will describe some of the more common kinds of monochrome (i.e., black and white) and color printers. Monochrome Printers The cheapest kind of printer is the matrix printer, in which a print head containing between 7 and 24 electromagnetically activatable needles is scanned across each print line. Low-end printers have seven needles, for printing, say, 80 characters in a 5 × 7 matrix across the line. In effect, the print line then consists of 7 horizontal lines, each consisting of 5 × 80 = 400 dots. Each dot can be printed or not printed, depending on the characters to be printed. Figure 2-34(a) illustrates the letter ‘‘A’’ printed on a 5 × 7 matrix.

(a)

(b)

Figure 2-34. (a) The letter ‘‘A’’ on a 5 × 7 matrix. (b) The letter ‘‘A’’ printed with 24 overlapping needles.

The print quality can be increased by two techniques: using more needles and having the circles overlap. Figure 2-34(b) shows an ‘‘A’’ printed using 24 needles that produce overlapping dots. Usually, multiple passes over each scan line are

SEC. 2.4

INPUT/OUTPUT

113

required to produce overlapping dots, so increased quality goes hand in hand with slower printing rates. Most matrix printers can operate in several modes, offering different trade-offs between print quality and speed. Matrix printers are cheap (especially in terms of consumables) and highly reliable, but slow, noisy, and poor at graphics. They have three main uses in current systems. First, they are popular for printing on large (> 30 cm) preprinted forms. Second, they are good at printing on small pieces of paper, such as cash register receipts, ATM machine or credit card transaction slips, or airline boarding passes. Third, for printing on multipart continuous forms with carbon paper embedded between the copies, they are usually the cheapest technology. For low-cost home printing, inkjet printers are a favorite. The movable print head, which holds an ink cartridge, is swept horizontally across the paper by a belt while ink is sprayed from its tiny nozzles. The ink droplets have a volume of about 1 picoliter, which means that 100 million of them would fit nicely in a single drop of water. Inkjet printers come in two varieties: piezoelectric (used by Epson) and thermal (used by Canon, HP, and Lexmark). The piezoelectric inkjet printers have a special kind of crystal next to the ink chamber. When a voltage is applied to the crystal, it deforms slightly, forcing a droplet of ink out of the nozzle. The higher the voltage, the larger the droplet, allowing the software to control the droplet size. Thermal inkjet printers (also called bubblejet printers) contain a tiny resistor inside each nozzle. When a voltage is applied to the resistor, it heats up extremely fast, instantly raising the temperature of the ink touching it to the boiling point until the ink vaporizes to form a gas bubble. The gas bubble takes up more volume than the ink that created it, producing pressure in the nozzle. The only place the ink can go is out the front of the nozzle onto the paper. The nozzle is then cooled and the resulting vacuum sucks in another ink droplet from the ink tank. The speed of the printer is limited by how fast the boil/cool cycle can be repeated. The droplets are all the same size, but usually smaller than what the piezo-electric printers produce. Inkjet printers typically have resolutions of at least 1200 dpi (dots per inch) and at the high end, 4800 dpi, They are cheap, quiet, and have good quality, although they are also slow, and use expensive ink cartridges. When the best of the high-end inkjet printers is used to print a high-resolution photograph on specially-coated photographic paper, the results are indistinguishable from conventional photography, even up to 8 × 10 prints. Probably the most exciting development in printing since Johann Gutenberg invented movable type in the fifteenth century is the laser printer. This device combines a high quality image, excellent flexibility, great speed, and moderate cost into a single peripheral. Laser printers use almost the same technology as photocopy machines. In fact, many companies make devices that combine copying and printing (and sometimes fax as well).

114

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

The basic technology is illustrated in Fig. 2-35. The heart of the printer is a rotating precision drum (or in some high-end systems, a belt). At the start of each page cycle, it is charged up to about 1000 volts and coated with a photosensitive material. Then light from a laser is scanned along the length of the drum much like the electron beam in a CRT, only instead of achieving the horizontal deflection using a voltage, a rotating octagonal mirror is used to scan the length of the drum. The light beam is modulated to produce a pattern of light and dark spots. The spots where the beam hits lose their electrical charge. Rotating octagonal mirror

Laser

Drum sprayed and charged Light beam strikes drum

Toner

Drum

Scraper Discharger Heated rollers

Blank paper

Stacked output

Figure 2-35. Operation of a laser printer.

After a line of dots has been painted, the drum rotates a fraction of a degree to allow the next line to be painted. Eventually, the first line of dots reaches the toner, a reservoir of an electrostatically sensitive black powder. The toner is attracted to those dots that are still charged, thus forming a visual image of that line. A little later in the transport path, the toner-coated drum is pressed against the paper, transferring the black powder to the paper. The paper is then passed through heated rollers to fuse the toner to the paper permanently, fixing the image. Later in its rotation, the drum is discharged and scraped clean of any residual toner, preparing it for being charged and coated again for the next page. That this process is an exceedingly complex combination of physics, chemistry, mechanical engineering, and optical engineering hardly needs to be said. Nevertheless, complete assemblies, called print engines, are available from several vendors. Laser printer manufacturers combine the print engines with their own electronics and software to make a complete printer. The electronics consists of a fast embedded CPU along with megabytes of memory to hold a full-page bit map and numerous fonts, some of them built in and some of them downloadable.

SEC. 2.4

115

INPUT/OUTPUT

Most printers accept commands that describe the pages to be printed (as opposed to simply accepting bit maps prepared by the main CPU). These commands are given in languages such as HP’s PCL and Adobe’s PostScript. Laser printers at 600-dpi and up can do a reasonable job of printing black and white photographs but the technology is trickier than it might at first appear. Consider a photograph scanned in at 600 dpi that is to be printed on a 600 dpi printer. The scanned image contains 600 × 600 pixels/inch, each one consisting of a gray value from 0 (white) to 255 (black). The printer can also print 600 dpi, but each printed pixel is either black (toner present) or white (no toner present). Gray values cannot be printed. The usual solution to printing images with gray values is to use halftoning, the same as commercially printed posters. The image is broken up into halftone cells, each typically 6 × 6 pixels. Each cell can contain between 0 and 36 black pixels. The eye perceives a cell with many pixels as darker than one with fewer pixels. Gray values in the range 0 to 255 are represented by dividing this range into 37 zones. Values from 0 to 6 are in zone 0, values from 7 to 13 are in zone 1, and so on (zone 36 is slightly smaller than the others because 37 does not divide 256 exactly). Whenever a gray value in zone 0 is encountered, its halftone cell on the paper is left blank, as illustrated in Fig. 2-36(a). A zone 1 value is printed as 1 black pixel. A zone 2 value is printed as 2 black pixels, as shown in Fig. 2-36(b). Other zone values are shown in Fig. 2-36(c)-(f). Of course, taking a photograph scanned at 600 dpi and halftoning this way reduces the effective resolution to 100 cells/inch, called the halftone screen frequency, conventionally measured in lpi (lines per inch).

(a)

(b)

(c)

(d)

(e)

(f)

Figure 2-36. Halftone dots for various gray scale ranges. (a) 0–6. (b) 14–20. (c) 28–34. (d) 56–62. (e) 105–111. (f) 161–167.

Color Printers Color images can be viewed in one of two ways: transmitted light and reflected light. Transmitted light images, such as those produced on CRT monitors, are built up from the linear superposition of the three additive primary colors, red, green, and blue. Reflected light images, such as color photographs and pictures in glossy magazines, absorb certain wavelengths of light and reflect the rest.

116

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

These are built up from a linear superposition of the three subtractive primary colors, cyan (all red absorbed), yellow (all blue absorbed), and magenta (all green absorbed). In theory, every color can be produced by mixing cyan, yellow, and magenta ink. In practice it is difficult to get the inks pure enough to absorb all light and produce a true black. For this reason, nearly all color printing systems use four inks: cyan, yellow, magenta, and black. These systems are called CYMK printers (K is for blacK, to avoid confusion with Blue). Monitors, in contrast, use transmitted light and the RGB system for producing colors. The complete set of colors that a display or printer can produce is called its gamut. No device has a gamut that matches the real world, since at best each color comes in 256 intensities, giving only 16,777,216 discrete colors. Imperfections in the technology reduce the total more, and the remaining ones are not always uniformly spaced over the color spectrum. Furthermore, color perception has a lot to do with how the rods and cones in the human retina work, and not just the physics of light. As a consequence of the above observations, converting a color image that looks fine on the screen to an identical printed one is far from trivial. Among the problems are 1. Color monitors use transmitted light; color printers use reflected light. 2. CRTs produce 256 intensities per color; color printers must halftone. 3. Monitors have a dark background; paper has a light background. 4. The RGB and CMYK gamuts are different. Getting printed color images to match real life (or even to match screen images) requires device calibration, sophisticated software, and considerable expertise on the part of the user. Five technologies are in common use for color printing, all of them based on the CMYK system. At the low end are color ink jet printers. They work the same way as monochrome ink jet printers, but with four cartridges (for C, M, Y, and K) instead of one. They give good results for color graphics and passable results for photographs at modest cost (the printers are cheap but the ink cartridges are not). For best results, special ink and paper should be used. Two kinds of ink exist. Dye-based inks consist of colored dyes dissolved in a fluid carrier. They give bright colors and flow easily. Their main disadvantage is that they fade when exposed to ultraviolet light, such as that contained in sunlight. Pigment-based ink contains solid particles of pigment suspended in a fluid carrier that evaporates from the paper, leaving the pigment behind. They do not fade in time but are not as bright as dye-based inks and the pigment particles have a tendency to clog the nozzles, requiring periodic cleaning. Coated or glossy paper is required for printing photographs. These kinds of paper have been specially designed to hold the ink droplets and not let them spread out.

SEC. 2.4

INPUT/OUTPUT

117

A step up from ink jet printers leads to the solid ink printers. These accept four solid blocks of a special waxy ink which are then melted into hot ink reservoirs. Startup times of these printers can be as much as 10 minutes, while the ink blocks are melting. The hot ink is sprayed onto the paper, where it solidifies and is fused with the paper by forcing it between two hard rollers The third kind of color printer is the color laser printer. It works like its monochrome cousin, except that separate C, Y, M, and K images are laid down and transferred to a roller using four different toners. Since the full bit map is generally produced in advance, a 1200 × 1200 dpi image for a page containing 80 square inches needs 115 million pixels. With 4 bits/pixels, the printer needs 55 MB just for the bit map, exclusive of memory for the internal processors, fonts, etc. This requirement makes color laser printers expensive, but printing is fast, the quality is high, and the images are stable over time. The fourth kind of color printer is the wax printer. It has a wide ribbon of four-color wax that is segmented into page-size bands. Thousands of heating elements melt the wax as the paper moves under it. The wax is fused to the paper in the form of pixels using the CMYK system. Wax printers used to be the main color printing technology, but they are being replaced by the other kinds, which have cheaper consumables. The fifth kind of color printer is the dye sublimation printer. Although it has Freudian undertones, sublimation is the scientific name for a solid changing into a gas without passing through the liquid state. Dry ice (frozen carbon dioxide) is a well-known material that sublimates. In a dye sublimation printer, a carrier containing the CMYK dyes passes over a thermal print head containing thousands of programmable heating elements. The dyes are vaporized instantly and absorbed by a special paper close by. Each heating element can produce 256 different temperatures. The higher the temperature, the more dye that is deposited and the more intense the color. Unlike all the other color printers, nearly continuous colors are possible for each pixel, so no halftoning is needed. Small snapshot printers often use the dye sublimation process to produce highly realistic photographic images on special (and expensive) paper.

2.4.5 Telecommunications Equipment Most computers nowadays are connected to a computer network, often the Internet. Achieving this access requires special equipment. In this section we will see how this equipment works. Modems With the growth of computer usage in the past years, it is common for one computer to need to communicate with another computer. For example, many people have personal computers at home that they use for communicating with

118

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

their computer at work, with an Internet Service Provider, or with a home banking system. In many cases, the telephone line provides the physical communication. However, a raw telephone line (or cable) is not suitable for transmitting computer signals, which generally represent a 0 as 0 volts and a 1 as 3 to 5 volts as shown in Fig. 2-37(a). Two-level signals suffer considerable distortion when transmitted over a voice-grade telephone line, thereby leading to transmission errors. A pure sine wave signal at a frequency of 1000 to 2000 Hz, called a carrier, can be transmitted with relatively little distortion, however, and this fact is exploited as the basis of most telecommunication systems.

(a)

(b)

Voltage

V2 V1

0

1

0

0

1

Time 0 1

1

0

0

0

1

0

0

High amplitude

Low amplitude

High frequency

Low frequency

(c)

(d)

Phase change

Figure 2-37. Transmission of the binary number 01001011000100 over a telephone line bit by bit. (a) Two-level signal. (b) Amplitude modulation. (c) Frequency modulation. (d) Phase modulation.

Because the pulsations of a sine wave are completely predictable, a pure sine wave transmits no information at all. However, by varying the amplitude, frequency, or phase, a sequence of 1s and 0s can be transmitted, as shown in Fig. 237. This process is called modulation. In amplitude modulation [see Fig. 237(b)], two different voltage levels are used, for 0 and 1, respectively. A person listening to digital data transmitted at a very low data rate would hear a loud noise for a 1 and no noise for a 0. In frequency modulation [see Fig. 2-37(c)], the voltage level is constant but the carrier frequency is different for 1 and 0. A person listening to frequency modulated digital data would hear two tones, corresponding to 0 and 1. Frequency modulation is often referred to as frequency shift keying.

SEC. 2.4

INPUT/OUTPUT

119

In simple phase modulation [see Fig. 2-37(d)], the amplitude and frequency do not change, but the phase of the carrier is reversed 180 degrees when the data switch from 0 to 1 or 1 to 0. In more sophisticated phase-modulated systems, at the start of each indivisible time interval, the phase of the carrier is abruptly shifted by 45, 135, 225, or 315 degrees, to allow 2 bits per time interval, called dibit phase encoding. For example, a phase shift of 45 degrees could represent 00, a phase shift of 135 degrees could represent 01, and so on. Other schemes, for transmitting 3 or more bits per time interval also exist. The number of time intervals (i.e., the number of potential signal changes per second) is baud rate. With 2 or more bits per interval, the bit rate will exceed the baud rate. Many people confuse these two terms. If the data to be transmitted consist of a series of 8-bit characters, it would be desirable to have a connection capable of transmitting 8 bits simultaneously—that is, eight pairs of wires. Because voice-grade telephone lines provide only one channel, the bits must be sent serially, one after another (or in groups of two if dibit encoding is being used). The device that accepts characters from a computer in the form of two-level signals, one bit at a time, and transmits the bits in groups of one or two, in amplitude-, frequency-, or phase-modulated form, is the modem. To mark the start and end of each character, an 8-bit character is normally sent preceded by a start bit and followed by a stop bit, making 10 bits in all. The transmitting modem sends the individual bits within one character at regularly-spaced time intervals. For example, 9600 baud implies one signal change every 104 µsec. A second modem at the receiving end is used to convert a modulated carrier to a binary number. Because the bits arrive at the receiver at regularly-spaced intervals, once the receiving modem has determined the start of the character, its clock tells it when to sample the line to read the incoming bits. Modern modems operate at data rates ranging from 28,800 bits/sec to 57,600 bits/sec, usually at much lower baud rates. They use a combination of techniques to send multiple bits per baud, modulating the amplitude, frequency, and phase. Nearly all of them are full-duplex, meaning they can transmit in both directions at the same time (using different frequencies). Modems or transmission lines that can only transmit in one direction at a time (like a single-track railroad that can handle north-bound trains or south-bound trains but not at the same time) are called half-duplex. Lines that can only transmit in one direction are simplex. Digital Subscriber Lines When the telephone industry finally got to 56 kbps, it patted itself on the back for a job well done. Meanwhile, the cable TV industry was offering speeds up to 10 Mbps on shared cables, and satellite companies were planning to offer upward of 50 Mbps. As Internet access became an increasingly important part of their business, the telcos (telephone companies) began to realize they needed a more competitive product than dialup lines. Their answer was to start offering a new

120

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

digital Internet access service. Services with more bandwidth than standard telephone service are sometimes called broadband, although the term really is more of a marketing concept than a specific technical concept. Initially, there were many overlapping offerings, all under the general name of xDSL (Digital Subscriber Line), for various x. Below we will discuss what is probably going to become the most popular of these services, ADSL (Asymmetric DSL). Since ADSL is still being developed and not all the standards are fully in place, some of the details given below may change in time, but the basic picture should remain valid. For more information about ADSL, see (Summers, 1999; and Vetter et al., 2000). The reason that modems are so slow is that telephones were invented for carrying the human voice and the entire system has been carefully optimized for this purpose. Data have always been stepchildren. The wire, called the local loop, from each subscriber to the telephone company’s office has traditionally been limited to about 3000 Hz by a filter in the telco office. It is this filter that limits the data rate. The actual bandwidth of the local loop depends on its length, but for typical distances of a few kilometers, 1.1 MHz is feasible. The most common approach to offering ADSL is illustrated in Fig. 2-38. In effect, what it does is remove the filter and divide the available 1.1 MHz spectrum on the local loop into 256 independent channels of 4312.5 Hz each. Channel 0 is used for POTS (Plain Old Telephone Service). Channels 1–5 are not used, to keep the voice signal and data signals from interfering with each other. Of the remaining 250 channels, one is used for upstream control and one is used for downstream control. The rest are available for user data. ADSL is like having 250 modems. Power

256 4-kHz Channels

0 Voice

25

1100 kHz Upstream

Downstream

Figure 2-38. Operation of ADSL.

In principle, each of the remaining channels can be used for a full-duplex data stream, but harmonics, crosstalk, and other effects keep practical systems well below the theoretical limit. It is up to the provider to determine how many channels are used for upstream and how many for downstream. A 50–50 mix of upstream and downstream is technically possible, but most providers allocate something like 80%–90% of the bandwidth to the downstream channel since most users download more data than they upload. This choice gives rise to the ‘‘A’’ in ADSL. A common split is 32 channels for upstream and the rest downstream.

SEC. 2.4

121

INPUT/OUTPUT

Within each channel the line quality is constantly monitored and the data rate adjusted continuously as needed, so different channels may have different data rates. The actual data are sent using a combination of amplitude and phase modulation with up to 15 bits per baud. With, for example, 224 downstream channels and 15 bits/baud at 4000 baud, the downstream bandwidth is 13.44 Mbps. In practice, the signal-to-noise ratio is never good enough to achieve this rate, but 4–8 Mbps is possible on short runs over high-quality loops. A typical ADSL arrangement is shown in Fig. 2-39. In this scheme, the user or a telephone company technician must install a NID (Network Interface Device) on the customer’s premises. This small plastic box marks the end of the telephone company’s property and the start of the customer’s property. Close to the NID (or sometimes combined with it) is a splitter, an analog filter that separates the 0-4000 Hz band used by POTS from the data. The POTS signal is routed to the existing telephone or fax machine, and the data signal is routed to an ADSL modem. The ADSL modem is actually a digital signal processor that has been set up to act as 250 modems operating in parallel at different frequencies. Since most current ADSL modems are external, the computer must be connected to it at high speed. Usually, this is done by putting an Ethernet card in the computer and operating a very short two-node Ethernet containing only the computer and ADSL modem. (Ethernet is a popular and inexpensive local area network standard.) Occasionally the USB port is used instead of Ethernet. In the future, internal ADSL modem cards will no doubt become available.

Voice switch

Telephone Codec Splitter

Telephone line

Splitter NID Computer

DSLAM

To ISP Telephone company end office

ADSL modem

Ethernet Customer premises

Figure 2-39. A typical ADSL equipment configuration.

At the other end of the wire, on the telco side, a corresponding splitter is installed. Here the voice portion of the signal is filtered out and sent to the normal

122

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

voice switch. The signal above 26 kHz is routed to a new kind of device called a DSLAM (Digital Subscriber Line Access Multiplexer), which contains the same kind of digital signal processor as the ADSL modem. Once the digital signal has been recovered into a bit stream, packets are formed and sent off to the ISP. Internet over Cable Many cable TV companies are now offering Internet access over their cables. Since the technology is quite different from ADSL, it is worth looking at briefly. The cable operator in each city has a main office and a large number of boxes full of electronics, called headends, spread all over its territory. The headends are connected to the main office by high-bandwidth cables or fiber optics. Each headend has one or more cables that run from it past hundreds of homes and offices. Each cable customer taps onto the cable as it passes the customer’s premises. Thus hundreds of users share the same cable to the headend. Usually, the cable has a bandwidth of about 750 MHz. This system is radically different from ADSL because each telephone user has a private (i.e., not shared) wire to the telco office. However, in practice, having your own 1.1 MHz channel to a telco office is not that different than sharing a 200-MHz piece of cable spectrum to the headend with 400 users, half of whom are not using it at any one instant. It does mean, however, that a cable Internet user will get much better service at 4 A.M. than at 4 P.M whereas ADSL service is constant all day long. People intent on getting optimal Internet over cable service might wish to consider moving to a rich neighborhood (houses far apart so fewer customers per cable) or a poor neighborhood (nobody can afford Internet service). Since the cable is a shared medium, determining who may send when and at which frequency is a big issue. To see how that works, we have to briefly describe how cable TV operates. Cable television channels in North America normally occupy the 54–550 MHz region (except for FM radio from 88 to 108 MHz). These channels are 6 MHz wide, including guard bands to prevent signal leakage between channels. In Europe the low end is usually 65 MHz and the channels are 6–8 MHz wide for the higher resolution required by PAL and SECAM but otherwise the allocation scheme is similar. The low part of the band is not used for television transmission. When introducing Internet over cable, the cable companies had two problems to solve: 1. How to add Internet access without interfering with TV programs. 2. How to have two-way traffic when amplifiers are inherently one way. The solutions chosen are as follows. Modern cables operate well above 550 MHz, often to 750 MHz or more. The upstream (i.e., user to headend) channels go in

SEC. 2.4

123

INPUT/OUTPUT

the 5–42 MHz band (slightly higher in Europe) and the downstream (i.e., headend to user) traffic uses the frequencies at the high end, as illustrated in Fig. 2-40. 5 42 54 88 108

Upstream frequencies

Upstream data

0

TV FM

550 TV

750 MHz Downstream data

Downstream frequencies

Figure 2-40. Frequency allocation in a typical cable TV system used for Internet access.

Note that since the television signals are all downstream, it is possible to use upstream amplifiers that work only in the 5–42 MHz region and downstream amplifiers that work only at 54 MHz and up, as shown in the figure. Thus, we get an asymmetry in the upstream and downstream bandwidths because more spectrum is available above television than below it. On the other hand, most of the traffic is likely to be downstream, so cable operators are not unhappy with this fact of life. As we saw earlier, telephone companies usually offer an asymmetric DSL service, even though they have no technical reason for doing so. Internet access requires a cable modem, a device that has two interfaces on it: one to the computer and one to the cable network. The computer-to-cable-modem interface is straightforward. It is normally Ethernet, just as with ADSL. In the future, the entire modem might be a small card plugged into the computer, just as with V.9x internal modems. The other end is more complicated. A large part of the cable standard deals with radio engineering, a subject far beyond the scope of this book. The only part worth mentioning here is that cable modems, like ADSL modems, are always on. They make a connection when turned on and maintain that connection as long as they are powered up because cable operators do not charge for connect time. To better understand how they work, let us see what happens when a cable modem is plugged in and powered up. The modem scans the downstream channels looking for a special packet periodically put out by the headend to provide system parameters to modems that have just come on-line. Upon finding this packet, the new modem announces its presence on one of the upstream channels. The headend responds by assigning the modem to its upstream and downstream channels. These assignments can be changed later if the headend deems it necessary to balance the load. The modem then determines its distance from the headend by sending it a special packet and seeing how long it takes to get the response. This process is called

124

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

ranging. It is important for the modem to know its distance to accommodate the way the upstream channels operate and to get the timing right. They are divided in time in minislots. Each upstream packet must fit in one or more consecutive minislots. The headend announces the start of a new round of minislots periodically, but the starting gun is not heard at all modems simultaneously due to the propagation time down the cable. By knowing how far it is from the headend, each modem can compute how long ago the first minislot really started. Minislot length is network dependent. A typical payload is 8 bytes. During initialization, the headend also assigns each modem to a minislot to use for requesting upstream bandwidth. As a rule, multiple modems will be assigned the same minislot, which leads to contention. When a computer wants to send a packet, it transfers the packet to the modem, which then requests the necessary number of minislots for it. If the request is accepted, the headend puts an acknowledgement on the downstream channel telling the modem which minislots have been reserved for its packet. The packet is then sent, starting in the minislot allocated to it. Additional packets can be requested using a field in the header. On the other hand, if there is contention for the request minislot, there will be no acknowledgement and the modem just waits a random time and tries again. After each successive failure, the randomization time is doubled to spread out the load when there is heavy traffic. The downstream channels are managed differently from the upstream channels. For one thing, there is only one sender (the headend) so there is no contention and no need for minislots, which is actually just time division statistical multiplexing. For another, the traffic downstream is usually much larger than upstream, so a fixed packet size of 204 bytes is used. Part of that is a ReedSolomon error-correcting code and some other overhead, leaving a user payload of 184 bytes. These numbers were chosen for compatibility with digital television using MPEG-2, so the TV and downstream data channels are formatted the same way. Logically, the connections are as depicted in Fig. 2-41. Getting back to modem initialization, once the modem has completed ranging and gotten its upstream channel, downstream channel, and minislot assignments, it is free to start sending packets. These packets go to the headend, which relays them over a dedicated channel to the cable company’s main office and then to the ISP (which may be the cable company itself). The first packet is one to the ISP requesting a network address (technically, an IP address), which is dynamically assigned. It also requests and gets an accurate time of day. The next step involves security. Since cable is a shared medium, anybody who wants to go to the trouble to do so can read all the traffic going past him. To prevent everyone from snooping on their neighbors (literally), all traffic is encrypted in both directions. Part of the initialization procedure involves establishing encryption keys. At first one might think that having two strangers, the headend and the modem, establish a secret key in broad daylight with thousands of people watching would be impossible to accomplish. Turns out it is not, but

SEC. 2.4

INPUT/OUTPUT

125

Figure 2-41. Typical details of the upstream and downstream channels in North America. QAM-64 (Quadrature Amplitude Modulation) allows 6 bits/Hz but only works at high frequencies. QPSK (Quadrature Phase Shift Keying) works at low frequencies but allows only 2 bits/Hz.

the technique used (the Diffie-Hellman algorithm) is beyond the scope of this book. See Kaufman et al. (2002) for a discussion of it. Finally, the modem has to log in and provide its unique identifier over the secure channel. At this point the initialization is complete. The user can now log in to the ISP and get to work. There is much more to be said about cable modems. Some relevant references are (Adams and Dulchinos, 2001; Donaldson and Jones, 2001; and DuttaRoy, 2001).

2.4.6 Digital Cameras An increasingly popular use of computers is for digital photography, making digital cameras a kind of computer peripheral. Let us briefly see how that works. All cameras have a lens that forms an image of the subject in the back of the camera. In a conventional camera, the back of the camera is lined with film, on which a latent image is formed when light strikes it. The latent image can be made visible by the action of certain chemicals in the film developer. A digital camera works the same way except that the film is replaced by a rectangular array of CCDs (Charge-Coupled Devices) that are sensitive to light. (Some digital cameras use CMOS, but we will concentrate on the more common CCDs here.) When light strikes a CCD, it acquires an electrical charge. The more light, the more charge. The charge can be read off by an analog-to-digital converter as an integer from 0 to 255 (on low-end cameras) or 0 to 4095 (on digital single lens reflex cameras). The basic arrangement is shown in Fig. 2-42. Each CCD produces a single value, independent of the color of light striking it. To form color images, the CCDs are organized in groups of four elements. A Bayer filter is placed on top of the CCD to allow only red light to strike one of

126

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

R GR G GB GB RGR G GB GB

Lens

One pixel is made up for four CCDs, one red, one blue, and two green

Diaphragm CPU RAM Flash memory or microdrive

CCD array

Digital camera Figure 2-42. A digital camera.

the four CCDs in each group, blue light to strike another one, and green light to strike the other two. Two greens are used because using four CCDs to represent one pixel is much more convenient than using three, and the eye is more sensitive to green light than to red or blue light. When a digital camera manufacturer claims a camera has, say, 6 million pixels, it is lying. The camera has 6 million CCDs, which together form 1.5 million pixels. The image will be read out as an array of 2828 × 2121 pixels (on low-end cameras) or 3000 times 2000 pixels (on digital SLRs), but the extra pixels are produced by interpolation by software inside the camera. When the camera’s shutter button is depressed, software in the camera performs three tasks: setting the focus, determining the exposure, and performing the white balance. The autofocus works by analyzing the high frequency information in the image and then moving the lens until it is maximized, to give the most detail. The exposure is determined by measuring the light falling on the CCDs and then adjusting the lens diaphragm and exposure time to have the light intensity fall in the middle of the CCDs’ range. Setting the white balance has to do with measuring the spectrum of the incident light to perform necessary color corrections later. Then the image is read off the CCDs and stored as a pixel array in the camera’s internal RAM. High-end digital SLRs used by photojournalists can shoot eight high-resolution frames per second for 5 seconds, and need around 1

SEC. 2.4

INPUT/OUTPUT

127

GB of internal RAM to store the images before processing and storing them permanently. Low-end cameras have less RAM, but still quite a bit. In the post-capture phase, the camera’s software applies the white balance color correction to compensate for reddish or bluish light (e.g., from a subject in shadow or use of a flash). Then it applies an algorithm to do noise reduction and another one to compensate for defective CCDs. After that, it attempts to sharpen the image (unless this feature has been disabled) by looking for edges and increasing the intensity gradient around them. Finally, the image may be compressed to reduce the amount of storage required. A common format is JPEG (Joint Photographic Experts Group), in which a two-dimensional spatial Fourier transform is applied and some of the high-frequency components omitted. The result of this transformation is that the image requires fewer bits to store but fine detail is lost. When all the in-camera processing is completed, the image is written to the storage medium, usually a flash memory or a tiny removable hard disk called a microdrive. The postprocessing and writing can take several seconds per image. When the user gets home, the camera can be connected to a computer, usually using, for example, a USB or FireWire cable. The images are then transferred from the camera to the computer’s hard disk. Using special software, such as Adobe Photoshop, the user can then crop the image, adjust brightness, contrast, and color balance, sharpen, blur or remove portions of the image, and apply numerous filters. When the user is content with the result, the image files can be printed on a color printer, uploaded over the Internet to a photofinisher, or written to CD-ROM or DVD for archival storage or subsequent printing. The amount of computing power, RAM, hard disk space, and software in a digital SLR camera is mind boggling. Not only does the computer have to do all the things mentioned above, but it also has to communicate with the CPU in the lens and the CPU in the flash, refresh the image on the LCD screen, and manage all the buttons, wheels, lights, displays, and gizmos on the camera in real time. This is an extremely powerful embedded system, often rivaling a desktop computer of only a few years earlier.

2.4.7 Character Codes Each computer has a set of characters that it uses. As a bare minimum, this set includes the 26 uppercase letters, the 26 lowercase letters, the digits 0 through 9, and a set of special symbols, such as space, period, minus sign, comma, and carriage return. In order to transfer these characters into the computer, each one is assigned a number: for example, a = 1, b = 2, ..., z = 26, + = 27, − = 28. The mapping of characters onto integers is called a character code. It is essential that communicating computers use the same code or they will not be able to understand one

128

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

another. For this reason, standards have been developed. Below we will examine two of the most important ones. ASCII One widely used code is called ASCII (American Standard Code for Information Interchange). Each ASCII character has 7 bits, allowing for 128 characters in all. Figure 2-43 shows the ASCII code. Codes 0 to 1F (hexadecimal) are control characters and do not print. Many of the ASCII control characters are intended for data transmission. For example, a message might consist of an SOH (Start of Header) character, a header, an STX (Start of Text) character, the text itself, an ETX (End of Text) character and then an EOT (End of Transmission) character. In practice, however, the messages sent over telephone lines and networks are formatted quite differently, so the ASCII transmission control characters are not used much any more. The ASCII printing characters are straightforward. They include the upper and lowercase letters, digits, punctuation marks and a few math symbols. UNICODE The computer industry grew up mostly in the U.S., which led to the ASCII character set. ASCII is fine for English but less fine for other languages. French needs accents (e.g., syste`me); German needs diacritical marks (e.g., fu¨r), and so on. Some European languages have a few letters not found in ASCII, such as the German β and the Danish o/. Some languages have entirely different alphabets (e.g., Russian and Arabic), and a few languages have no alphabet at all (e.g., Chinese). As computers spread to the four corners of the globe, and software vendors want to sell products in countries where most users do not speak English, a different character set is needed. The first attempt at extending ASCII was IS 646, which added another 128 characters to ASCII, making it an 8-bit code called Latin-1. The additional characters were mostly Latin letters with accents and diacritical marks. The next attempt was IS 8859, which introduced the concept of a code page, a set of 256 characters for a particular language or group of languages. IS 8859-1 is Latin-1. IS 8859-2 handles the Latin-based Slavic languages (e.g., Czech, Polish, and Hungarian). IS 8859-3 contains the characters needed for Turkish, Maltese, Esperanto, and Galician, and so on. The trouble with the code page approach is that the software has to keep track of which page it is on, it is impossible to mix languages over pages, and the scheme does not cover Japanese and Chinese at all. A group of computer companies decided to solve this problem by forming a consortium to create a new system, called UNICODE, and getting it proclaimed an International Standard (IS 10646). UNICODE is now supported by some

SEC. 2.4

INPUT/OUTPUT

129

################################################################################# ! Hex ! Hex Name Meaning ! Name Meaning ################################################################################# ! ! ! NUL Null DLE Data Link Escape ! 0 ! 10 ! ! 1 ! 11 ! SOH Start Of Heading DC1 Device Control 1 ! 2 ! 12 ! STX Start Of Text DC2 Device Control 2 ! ! ! ETX End Of Text DC3 Device Control 3 ! 3 ! 13 ! ! 4 ! EOT End Of Transmission ! 14 DC4 Device Control 4 ! 5 ! 15 ! ENQ Enquiry NAK Negative AcKnowledgement ! ! ! ACK ACKnowledgement ! 16 SYN SYNchronous idle ! 6 ! BEL BELl ETB End of Transmission Block ! 7 ! 17 ! ! 8 ! 18 ! BS BackSpace CAN CANcel ! 9 ! ! HT Horizontal Tab 19 EM End of Medium ! ! ! LF Line Feed SUB SUBstitute ! A ! 1A ! ! B ! 1B ! VT Vertical Tab ESC ESCape ! C ! 1C ! FF Form Feed FS File Separator ! ! ! Return Separator Group Carriage GS 1D D CR ! ! ! SO Shift Out RS Record Separator ! 1E ! ! E ! !! !################################################################################# F SI Shift In 1F US Unit Separator ! ! ################################################################################# ! Hex Char ! Hex Char ! Hex Char ! Hex Char ! Hex Char ! Hex Char ! !################################################################################# ! ! ! ! ! ! (Space) ! 30 0 ! 40 @ ! 50 P ! 60 ‘ ! 70 p ! ! 20 ! 21 ! 31 ! 1 ! 41 A ! 51 Q ! 61 a ! 71 q ! ! 22 ! ! ! ! ! " 32 2 42 B 52 R 62 b 72 r ! ! ! ! ! ! ! ! 23 # 33 3 43 C 53 S 63 c 73 s ! ! ! ! ! ! ! ! 24 ! 34 $ 4 ! 44 D ! 54 T ! 64 d ! 74 t ! ! 25 ! 35 % 5 ! 45 E ! 55 U ! 65 e ! 75 u ! ! ! ! ! ! ! ! 26 & 36 6 46 F 56 V 66 f 76 v ! ! ! ! ! ! ! ’ 7 ! 47 G ! 57 W ! 67 g ! 77 w ! ! 27 ! 37 ! 28 ! 38 ( 8 ! 48 H ! 58 X ! 68 h ! 78 x ! ! 29 ! 39 ) 9 ! 49 I ! 59 Y ! 69 i ! 79 y ! ! ! ! ! ! ! ! * : ! 4A J ! 5A Z ! 6A j ! 7A z ! ! 2A ! 3A ! 2B ! 3B + ; ! 4B K ! 5B [ ! 6B k ! 7B { ! ! 2C ! 3C , < ! 4C L ! 5C \ ! 6C l ! 7C | ! ! ! ! ! ! ! ! 2D 3D = 4D M 5D ] 6D m 7D } ! ! ! ! ! ! ! . > ! 4E N ! 5E ˆ ! 6E n ! 7E ~ ! ! 2E ! 3E !!################################################################################# !! 3F 2F / ? !! 4F O !! 5F # !! 6F o !! 7F DEL !! Figure 2-43. The ASCII character set.

programming languages (e.g., Java), some operating systems (e.g., Windows XP), and many applications. It is likely to become increasingly accepted as the computer industry goes global. The idea behind UNICODE is to assign every character and symbol a unique 16-bit value, called a code point. No multibyte characters or escape sequences are used. Having every symbol be 16 bits makes writing software simpler.

130

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

With 16-bit symbols, UNICODE has 65,536 code points. Since the world’s languages collectively use about 200,000 symbols, code points are a scarce resource that must be allocated with great care. About half the code points have already been allocated, and the UNICODE consortium is continually reviewing proposals to eat up the rest. To speed the acceptance of UNICODE, the consortium cleverly used Latin-1 as code points 0 to 255, making conversion between ASCII and UNICODE easy. To avoid wasting code points, each diacritical mark has its own code point. It is up to software to combine diacritical marks with their neighbors to form new characters. The code point space is divided up into blocks, each one a multiple of 16 code points. Each major alphabet in UNICODE has a sequence of consecutive zones. Some examples (and the number of code points allocated) are Latin (336), Greek (144), Cyrillic (256), Armenian (96), Hebrew (112), Devanagari (128), Gurmukhi (128), Oriya (128), Telugu (128), and Kannada (128). Note that each of these languages has been allocated more code points than it has letters. This choice was made in part because many languages have multiple forms for each letter. For example, each letter in English has two forms—lowercase and UPPERCASE. Some languages have three or more forms, possibly depending on whether the letter is at the start, middle, or end of a word. In addition to these alphabets, code points have been allocated for diacritical marks (112), punctuation marks (112), subscripts and superscripts (48), currency symbols (48), math symbols (256), geometric shapes (96), and dingbats (192). After these come the symbols needed for Chinese, Japanese, and Korean. First are 1024 phonetic symbols (e.g., katakana and bopomofo) and then the unified Han ideographs (20,992) used in Chinese and Japanese, and the Korean Hangul syllables (11,156). To allow users to invent special characters for special purposes, 6400 code points have been allocated for local use. While UNICODE solves many problems associated with internationalization, it does not (attempt to) solve all the world’s problems. For example, while the Latin alphabet is in order, the Han ideographs are not in dictionary order. As a consequence, an English program can examine ‘‘cat’’ and ‘‘dog’’ and sort them alphabetically by simply comparing the UNICODE value of their first character. A Japanese program needs external tables to figure out which of two symbols comes before the other in the dictionary. Another issue is that new words are popping up all the time. Fifty years ago nobody talked about applets, cyberspace, gigabytes, lasers, modems, smileys, or videotapes. Adding new words in English does not require new code points. Adding them in Japanese does. In addition to new technical words, there is a demand for adding at least 20,000 new (mostly Chinese) personal and place names. Blind people think Braille should be in there, and special interest groups of all kinds want what they perceive as their rightful code points. The UNICODE consortium reviews and decides on all new proposals.

SEC. 2.4

INPUT/OUTPUT

131

UNICODE uses the same code point for characters that look almost identical but have different meanings or are written slightly differently in Japanese and Chinese (as though English word processors always spelled ‘‘blue’’ as blew’’ because they sound the same). Some people view this as an optimization to save scarce code points; others see it as Anglo-Saxon cultural imperialism (and you thought assigning 16-bit values to characters was not highly political?). To make matters worse, a full Japanese dictionary has 50,000 kanji (excluding names), so with only 20,992 code points available for the Han ideographs, choices had to be made. Not all Japanese people think that a consortium of computer companies, even if a few of them are Japanese, is the ideal forum to make these choices.

2.5 SUMMARY Computer systems are built up from three types of components: processors, memories, and I/O devices. The task of a processor is to fetch instructions one at a time from a memory, decode them, and execute them. The fetch-decodeexecute cycle can always be described as an algorithm and, in fact, is sometimes carried out by a software interpreter running at a lower level. To gain speed, many computers now have one or more pipelines or have a superscalar design with multiple functional units that operate in parallel. Systems with multiple processors are increasingly common. Parallel computers include array processors, on which the same operation is performed on multiple data sets at the same time, multiprocessors, in which multiple CPUs share a common memory, and multicomputers, in which multiple computers each have their own memories but communicate by message passing. Memories can be categorized as primary or secondary. The primary memory is used to hold the program currently being executed. Its access time is short—a few tens of nanoseconds at most—and independent of the address being accessed. Caches reduce this access time even more. Some memories are equipped with error-correcting codes to enhance reliability. Secondary memories, in contrast, have access times that are much longer (milliseconds or more) and dependent on the location of the data being read or written. Tapes, magnetic disks and optical disks are the most common secondary memories. Magnetic disks come in many varieties, including floppy disks, Winchester disks, IDE disks, SCSI disks, and RAIDs. Optical disks include CDROMs, CD-Rs, and DVDs. I/O devices are used to transfer information into and out of the computer. They are connected to the processor and memory by one or more buses. Examples are terminals, mice, printers, and modems. Most I/O devices use the ASCII character code, although UNICODE is rapidly gaining acceptance as the computer industry goes global.

132

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

PROBLEMS 1. Consider the operation of a machine with the data path of Figure 2-2. Suppose that loading the ALU input registers takes 5 nsec, running the ALU takes 10 nsec, and storing the result back in the register scratchpad takes 5 nsec. What is the maximum number of MIPS this machine is capable of in the absence of pipelining? 2. What is the purpose of step 2 in the list of Sec. 2.1.2? What would happen if this step were omitted? 3. On computer 1, all instructions take 10 nsec to execute. On computer 2, they all take 5 nsec to execute. Can you say for certain that computer 2 is faster? Discuss. 4. Imagine you are designing a single-chip computer for an embedded system. The chip is going to have all its memory on chip and running at the same speed as the CPU with no access penalty. Examine each of the principles discussed in Sec. 2.1.4 and tell whether they are so important (assuming that high performance is still desired). 5. A certain computation is highly sequential—that is, each step depends on the one preceding it. Would an array processor or a pipeline processor be more appropriate for this computation? Explain. 6. To compete with the newly-invented printing press, a medieval monastery decided to mass-produce handwritten paperback books by assembling a vast number of scribes in a huge hall. The head monk would then call out the first word of the book to be produced and all the scribes would copy it down. Then the head monk would call out the second word and all the scribes would copy it down. This process was repeated until the entire book had been read aloud and copied. Which of the parallel processor systems discussed in Sec. 2.1.6 does this system resemble most closely? 7. As one goes down the five-level memory hierarchy discussed in the text, the access time increases. Make a reasonable guess about the ratio of the access time of optical disk to that of register memory. Assume that the disk is already on-line. 8. Sociologists can get three possible answers to a typical survey question such as ‘‘Do you believe in the tooth fairy?’’—namely, yes, no, and no opinion. With this in mind, the Sociomagnetic Computer Company has decided to build a computer to process survey data. This computer has a trinary memory—that is, each byte (tryte?) consists of 8 trits, with a trit holding a 0, 1, or 2. How many trits are needed to hold a 6-bit number? Give an expression for the number of trits needed to hold n bits. 9. Compute the data rate of the human eye using the following information. The visual field consists of about 106 elements (pixels). Each pixel can be reduced to a superposition of the three primary colors, each of which has 64 intensities. The time resolution is 100 msec. 10. Compute the data rate of the human ear from the following information. People can hear frequencies up to 22 kHz. To capture all the information in a sound signal at 22 kHz, it is necessary to sample the sound at twice that frequency, that is, at 44 kHz. A 16-bit sample is probably enough to capture most of the auditory information (i.e., the ear cannot distinguish more than 65,535 intensity levels).

CHAP. 2

PROBLEMS

133

11. Genetic information in all living things is coded as DNA molecules. A DNA molecule is a linear sequence of the four basic nucleotides: A, C, G, and T. The human genome contains approximately 3 × 109 nucleotides in the form of about 30,000 genes. What is the total information capacity (in bits) of the human genome? What is the maximum information capacity (in bits) of the average gene? 12. A certain computer can be equipped with 268,435,456 bytes of memory. Why would a manufacturer choose such a peculiar number, instead of an easy-to-remember number like 250,000,000? 13. Devise a 7-bit even-parity Hamming code for the digits 0 to 9. 14. Devise a code for the digits 0 to 9 whose Hamming distance is 2. 15. In a Hamming code, some bits are ‘‘wasted’’ in the sense that they are used for checking and not information. What is the percentage of wasted bits for messages whose total length (data + check bits) is 2n − 1? Evaluate this expression numerically for values of n from 3 to 10. 16. The disk illustrated in Figure 2-19 has 1024 sectors/track and a rotation rate of 7200 RPM. What is the sustained transfer rate of the disk over one track? 17. A computer has a bus with a 5 nsec cycle time, during which it can read or write a 32-bit word from memory. The computer has an Ultra4-SCSI disk that uses the bus and runs at 160 Mbytes/sec. The CPU normally fetches and executes one 32-bit instruction every 1 nsec. How much does the disk slow down the CPU? 18. Imagine you are writing the disk management part of an operating system. Logically, you represent the disk as a sequence of blocks, from 0 on the inside to some maximum on the outside. As files are created, you have to allocate free sectors. You could do it from the outside in or the inside out. Does it matter which strategy you choose? Explain your answer. 19. How long does it take to read a disk with 10,000 cylinders, each containing four tracks of 2048 sectors? First, all the sectors of track 0 are to be read starting at sector 0, then all the sectors of track 1 starting at sector 0, and so on. The rotation time is 10 msec, and a seek takes 1 msec between adjacent cylinders and 20 msec for the worst case. Switching between tracks of a cylinder can be done instantaneously. 20. RAID level 3 is able to correct single-bit errors using only one parity drive. What is the point of RAID level 2? After all, it also can only correct one error and takes more drives to do so. 21. What is the exact data capacity (in bytes) of a mode 2 CD-ROM containing the nowstandard 80-min media? What is the capacity for user data in mode 1? 22. To burn a CD-R, the laser must pulse on and off at a high speed. When running at 10x speed in mode 1, what is the pulse length, in nanoseconds? 23. To be able to fit 133 minutes worth of video on a single-sided single-layer DVD, a fair amount of compression is required. Calculate the compression factor required. Assume that 3.5 GB of space is available for the video track, that the image resolution is 720 × 480 pixels with 24-bit color, and images are displayed at 30 frames/sec.

134

COMPUTER SYSTEMS ORGANIZATION

CHAP. 2

24. Blu-Ray runs at 4.5 MB/sec and has a capacity of 25 GB. How long does it take to read the entire disk? 25. The transfer rate between a CPU and its associated memory is orders of magnitude higher than the mechanical I/O transfer rate. How can this imbalance cause inefficiencies? How can it be alleviated? 26. A manufacturer advertises that its color bit-map terminal can display 224 different colors. Yet the hardware only has 1 byte for each pixel. How can this be done? 27. A bit-map terminal has a 1600 × 1200 display. The display is redrawn 75 times a second. How long is the pulse corresponding to one pixel? 28. In a certain font, a monochrome laser printer can print 50 lines of 80 characters per page. The average character occupies a box 2 mm × 2 mm, about 25% of which is toner. The rest is blank. The toner layer is 25 microns thick. The printer’s toner cartridge measures 25 × 8 × 2 cm. How many pages is one toner cartridge good for? 29. When odd-parity ASCII text is transmitted asynchronously at a rate of 5600 characters/sec over a 56,000 bps modem, what percent of the received bits actually contain data (as opposed to overhead)? 30. The Hi-Fi Modem Company has just designed a new frequency-modulation modem that uses 64 frequencies instead of just 2. Each second is divided into n equal time intervals, each of which contains one of the 64 possible tones. How many bits per second can this modem transmit, using synchronous transmission? 31. An Internet user has subscribed to a 2 Mbps ADSL service. Her neighbor has subscribed to a cable Internet service that has a shared bandwidth of 12 MHz. The modulation scheme in use is QAM-64. There are n houses on the cable, each with one computer. A fraction f of these computers are online at any one time. Under what conditions will the cable user get better service than the ADSL user? 32. A digital camera has a resolution of 3000 × 2000 pixels, with 3 bytes/pixel for RGB color. The manufacturer of the camera wants to be able to write a JPEG image at a 5x compression factor to the flash memory in 2 sec. What data rate is required? 33. A high-end digital camera has a sensor with 16 million pixels, each with 3 bytes/pixel. How many pictures can be stored on a 1-GB flash memory card if the compression factor is 5x? Assume that 1 GB means 230 bytes. 34. Estimate how many characters, including spaces, a typical computer science textbook contains. How many bits are needed to encode a book in ASCII with parity? How many CD-ROMs are needed to store a computer science library of 10,000 books? How many double-side, dual-layer DVDs are needed for the same library? 35. Write a procedure hamming(ascii, encoded ) that converts the low-order 7 bits of ascii into an 11-bit integer codeword stored in encoded. 36. Write a function distance(code, n, k) that takes an array code of n characters of k bits each as input, and returns the distance of the character set as output.

3 THE DIGITAL LOGIC LEVEL

At the bottom of the hierarchy of Fig. 1-2 we find the digital logic level, the computer’s real hardware. In this chapter, we will examine many aspects of digital logic, as a building block for the study of higher levels in subsequent chapters. This subject is on the boundary of computer science and electrical engineering, but the material is self-contained, so no previous hardware or engineering experience is needed to follow it. The basic elements from which all digital computers are constructed are amazingly simple. We will begin our study by looking at these basic elements and also at the special two-valued algebra (Boolean algebra) used to analyze them. Next we will examine some fundamental circuits that can be built using gates in simple combinations, including circuits for doing arithmetic. The following topic is how gates can be combined to store information, that is, how memories are organized. After that, we come to the subject of CPUs and especially how single-chip CPUs interface with memory and peripheral devices. Numerous examples from industry will be discussed later in this chapter.

3.1 GATES AND BOOLEAN ALGEBRA Digital circuits can be constructed from a small number of primitive elements by combining them in innumerable ways. In the following sections we will describe these primitive elements, show how they can be combined, and introduce a powerful mathematical technique that can be used to analyze their behavior. 135

136

THE DIGITAL LOGIC LEVEL

CHAP. 3

3.1.1 Gates A digital circuit is one in which only two logical values are present. Typically, a signal between 0 and 1 volt represents one value (e.g., binary 0) and a signal between 2 and 5 volts represents the other value (e.g., binary 1). Voltages outside these two ranges are not permitted. Tiny electronic devices, called gates, can compute various functions of these two-valued signals. These gates form the hardware basis on which all digital computers are built. The details of how gates work inside is beyond the scope of this book, belonging to the device level, which is below our level 0. Nevertheless, we will now digress ever so briefly to take a quick look at the basic idea, which is not difficult. All modern digital logic ultimately rests on the fact that a transistor can be made to operate as a very fast binary switch. In Fig. 3-1(a) we have shown a bipolar transistor (the circle) embedded in a simple circuit. This transistor has three connections to the outside world: the collector, the base, and the emitter. When the input voltage, Vin , is below a certain critical value, the transistor turns off and acts like an infinite resistance. This causes the output of the circuit, Vout , to take on a value close to Vcc , an externally regulated voltage, typically +5 volts for this type of transistor. When Vin exceeds the critical value, the transistor switches on and acts like a wire, causing Vout to be pulled down to ground (by convention, 0 volts). +VCC +VCC +VCC Vout V1

Collector

Vout

Vout V2

Vin

V1

V2

Emitter

Base (a)

(b)

(c)

Figure 3-1. (a) A transistor inverter. (b) A NAND gate. (c) A NOR gate.

The important thing to notice is that when Vin is low, Vout is high, and vice versa. This circuit is thus an inverter, converting a logical 0 to a logical 1, and a logical 1 to a logical 0. The resistor (the jagged line) is needed to limit the amount of current drawn by the transistor so it does not burn out. The time required to switch from one state to the other is typically a few nanoseconds.

SEC. 3.1

137

GATES AND BOOLEAN ALGEBRA

In Fig. 3-1(b) two transistors are cascaded in series. If both V 1 and V 2 are high, both transistors will conduct and Vout will be pulled low. If either input is low, the corresponding transistor will turn off, and the output will be high. In other words, Vout will be low if and only if both V 1 and V 2 are high. In Fig. 3-1(c) the two transistors are wired in parallel instead of in series. In this configuration, if either input is high, the corresponding transistor will turn on and pull the output down to ground. If both inputs are low, the output will remain high. These three circuits, or their equivalents, form the three simplest gates. They are called NOT, NAND, and NOR gates, respectively. NOT gates are often called inverters; we will use the two terms interchangeably. If we now adopt the convention that ‘‘high’’ (Vcc volts) is a logical 1, and that ‘‘low’’ (ground) is a logical 0, we can express the output value as a function of the input values. The symbols used to depict these three gates are shown in Fig. 3-2(a)–(c), along with the functional behavior for each circuit. In these figures, A and B are inputs and X is the output. Each row specifies the output for a different combination of the inputs. NOT A

X

A

NAND X

(a)

X 1 0

X

B

B A 0 1

NOR

A

A 0 0 1 1

B 0 1 0 1 (b)

X 1 1 1 0

AND

A

X

B A 0 0 1 1

B 0 1 0 1 (c)

X 1 0 0 0

OR

A

X

B A 0 0 1 1

B 0 1 0 1

X 0 0 0 1

A 0 0 1 1

(d)

B 0 1 0 1

X 0 1 1 1

(e)

Figure 3-2. The symbols and functional behavior for the five basic gates.

If the output signal of Fig. 3-1(b) is fed into an inverter circuit, we get another circuit with precisely the inverse of the NAND gate—namely, a circuit whose output is 1 if and only if both inputs are 1. Such a circuit is called an AND gate; its symbol and functional description are given in Fig. 3-2(d). Similarly, the NOR gate can be connected to an inverter to yield a circuit whose output is 1 if either or both inputs is a 1 but 0 if both inputs are 0. The symbol and functional description of this circuit, called an OR gate, are given in Fig. 3-2(e). The small circles used as part of the symbols for the inverter, NAND gate, and NOR gate are called inversion bubbles. They are often used in other contexts as well to indicate an inverted signal. The five gates of Fig. 3-2 are the principal building blocks of the digital logic level. From the foregoing discussion, it should be clear that the NAND and NOR gates require two transistors each, whereas the AND and OR gates require three

138

THE DIGITAL LOGIC LEVEL

CHAP. 3

each. For this reason, many computers are based on NAND and NOR gates rather than the more familiar AND and OR gates. (In practice, all the gates are implemented somewhat differently, but NAND and NOR are still simpler than AND and OR.) In passing it is worth noting that gates may well have more than two inputs. In principle, a NAND gate, for example, may have arbitrarily many inputs, but in practice more than eight inputs is unusual. Although the subject of how gates are constructed belongs to the device level, we would like to mention the major families of manufacturing technology because they are referred to frequently. The two major technologies are bipolar and MOS (Metal Oxide Semiconductor). The major bipolar types are TTL (TransistorTransistor Logic), which had been the workhorse of digital electronics for years, and ECL (Emitter-Coupled Logic), which was used when very high-speed operation is required. For computer circuits, MOS has now largely taken over. MOS gates are slower than TTL and ECL but require much less power and take up much less space, so large numbers of them can be packed together tightly. MOS comes in many varieties, including PMOS, NMOS, and CMOS. While MOS transistors are constructed differently from bipolar transistors, their ability to function as electronic switches is the same. Most modern CPUs and memories use CMOS technology, which runs on +3.3 volts. This is all we will say about the device level. Readers interested in pursuing their study of this level should consult the suggested readings given in Chap. 9.

3.1.2 Boolean Algebra To describe the circuits that can be built by combining gates, a new type of algebra is needed, one in which variables and functions can take on only the values 0 and 1. Such an algebra is called a Boolean algebra, after its discoverer, the English mathematician George Boole (1815–1864). Strictly speaking, we are really referring to a specific type of Boolean algebra, a switching algebra, but the term ‘‘Boolean algebra’’ is so widely used to mean ‘‘switching algebra’’ that we will not make the distinction. Just as there are functions in ‘‘ordinary’’ (i.e., high school) algebra, so are there functions in Boolean algebra. A Boolean function has one or more input variables and yields a result that depends only on the values of these variables. A simple function, f, can be defined by saying that f(A) is 1 if A is 0 and f(A) is 0 if A is 1. This function is the NOT function of Fig. 3-2(a). Because a Boolean function of n variables has only 2n possible combinations of input values, the function can be completely described by giving a table with 2n rows, each row telling the value of the function for a different combination of input values. Such a table is called a truth table. The tables of Fig. 3-2 are all examples of truth tables. If we agree to always list the rows of a truth table in numerical order (base 2), that is, for two variables in the order 00, 01, 10, and 11, the function can be completely described by the 2n -bit binary number obtained by

SEC. 3.1

139

GATES AND BOOLEAN ALGEBRA

reading the result column of the truth table vertically. Thus NAND is 1110, NOR is 1000, AND is 0001, and OR is 0111. Obviously, only 16 Boolean functions of two variables exist, corresponding to the 16 possible 4-bit result strings. In contrast, ordinary algebra has an infinite number of functions of two variables, none of which can be described by giving a table of outputs for all possible inputs because each variable can take on any one of an infinite number of possible values. Figure 3-3(a) shows the truth table for a Boolean function of three variables: M = f (A, B, C). This function is the majority logic function, that is, it is 0 if a majority of its inputs are 0 and 1 if a majority of its inputs are 1. Although any Boolean function can be fully specified by giving its truth table, as the number of variables increases, this notation becomes increasingly cumbersome. Instead, another notation is frequently used. A B C

A B C

A 1 A 4

5

B

ABC

ABC

2 A 0 0 0 0 1 1 1 1

B 0 0 1 1 0 0 1 1

C 0 1 0 1 0 1 0 1

(a)

M 0 0 0 1 0 1 1 1

8

B 6

ABC

C 3 C

7

ABC

(b)

Figure 3-3. (a) The truth table for the majority function of three variables. (b) A circuit for (a).

To see how this other notation comes about, note that any Boolean function can be specified by telling which combinations of input variables give an output value of 1. For the function of Fig. 3-3(a) there are four combinations of input

M

140

THE DIGITAL LOGIC LEVEL

CHAP. 3

variables that make M 1. By convention, we will place a bar over an input variable to indicate that its value is inverted. The absence of a bar means that it is not inverted. Furthermore, we will use implied multiplication or a dot to mean the Boolean " AND function and + to mean the Boolean OR function. Thus, for"exam" ple, ABC takes the value 1 only when A = 1 and B = 0 and C = 1. Also, AB + BC is 1 only when (A = 1 and B = 0) or (B"= 1 and" C = 0)." The four rows of Fig. 33(a) producing 1 bits in the output are: ABC, ABC, ABC, and ABC. The function, M, is true (i.e., 1) if any one of these four conditions is true; hence we can write " " " M = ABC + ABC + ABC + ABC as a compact way of giving the truth table. A function of n variables can thus be described by giving a ‘‘sum’’ of at most 2n n-variable ‘‘product’’ terms. This formulation is especially important, as we will see shortly, because it leads directly to an implementation of the function using standard gates. It is important to keep in mind the distinction between an abstract Boolean function and its implementation by an electronic circuit. A Boolean function consists of variables, such as A, B, and C, and Boolean operators such as AND, OR, and NOT. A Boolean function is described by giving a truth table or a Boolean function such as " " F = ABC + ABC A Boolean function can be implemented by an electronic circuit (often in many different ways) using signals that represent the input and output variables and gates such as AND, OR, and NOT. We will generally use the notation AND, OR, and NOT when referring to the Boolean operators and AND, OR, and NOT when referring to the gates, but often it is ambiguous.

3.1.3 Implementation of Boolean Functions As mentioned above, the formulation of a Boolean function as a sum of up to 2n product terms leads directly to a possible implementation. Using Fig. 3-3 as an example, we can see how this implementation is accomplished. In Fig. 3-3(b), the inputs, A, B, and C, are shown at the left edge and the output function, M, is shown at the right edge. Because complements (inverses) of the input variables are needed, they are generated by tapping the inputs and passing them through the inverters labeled 1, 2, and 3. To keep the figure from becoming cluttered, we have drawn in six vertical lines, three of which are connected to the input variables, and three of which are connected to their complements. These lines provide a convenient source for the inputs to subsequent gates. For example, gates 5, 6, and 7 all use A as an input. In an actual circuit these gates would probably be wired directly to A without using any intermediate ‘‘vertical’’ wires. The circuit contains four AND gates, one for each term in the equation for M (i.e., one for each row in the truth table having a 1 bit in the result column). Each

SEC. 3.1

GATES AND BOOLEAN ALGEBRA

141

AND gate computes one row of the truth table, as indicated. Finally, all the product terms are ORed together to get the final result. The circuit of Fig. 3-3(b) uses a convention that we will use repeatedly throughout this book: when two lines cross, no connection is implied unless a heavy dot is present at the intersection. For example, the output of gate 3 crosses " all six vertical lines but it is connected only to C. Be warned that some authors use other conventions. From the example of Fig. 3-3 it should be clear how to implement a circuit for any Boolean function:

1. Write down the truth table for the function. 2. Provide inverters to generate the complement of each input. 3. Draw an AND gate for each term with a 1 in the result column. 4. Wire the AND gates to the appropriate inputs. 5. Feed the output of all the AND gates into an OR gate. Although we have shown how any Boolean function can be implemented using NOT, AND, and OR gates, it is often convenient to implement circuits using only a single type of gate. Fortunately, it is straightforward to convert circuits generated by the preceding algorithm to pure NAND or pure NOR form. To make such a conversion, all we need is a way to implement NOT, AND, and OR using a single gate type. The top row of Fig. 3-4 shows how all three of these can be implemented using only NAND gates; the bottom row shows how it can be done using only NOR gates. (These are straightforward, but there are other ways, too.) One way to implement a Boolean function using only NAND or only NOR gates is first follow the procedure given above for constructing it with NOT, AND, and OR. Then replace the multi-input gates with equivalent circuits using two-input gates. For example, A + B + C + D can be computed as (A + B) + (C + D), using three two-input OR gates. Finally, the NOT, AND, and OR gates are replaced by the circuits of Fig. 3-4. Although this procedure does not lead to the optimal circuits, in the sense of the minimum number of gates, it does show that a solution is always feasible. Both NAND and NOR gates are said to be complete, because any Boolean function can be computed using either of them. No other gate has this property, which is another reason they are often preferred for the building blocks of circuits.

3.1.4 Circuit Equivalence Circuit designers often try to reduce the number of gates in their products to reduce component cost, printed circuit board space, power consumption, and so on. To reduce the complexity of a circuit, the designer must find another circuit

142

THE DIGITAL LOGIC LEVEL

CHAP. 3

A

A

A

A (a)

A

AB

A

A+B

B B

A

AB

A

A+B

B

B (b)

(c)

Figure 3-4. Construction of (a) NOT, (b) AND, and (c) OR gates using only NAND gates or only NOR gates.

that computes the same function as the original but does so with fewer gates (or perhaps with simpler gates, for example, two-input gates instead of four-input gates). In the search for equivalent circuits, Boolean algebra can be a valuable tool. As an example of how Boolean algebra can be used, consider the circuit and truth table for AB + AC shown in Fig. 3-5(a). Although we have not discussed them yet, many of the rules of ordinary algebra also hold for Boolean algebra. In particular, AB + AC can be factored into A(B + C) using the distributive law. Figure 3-5(b) shows the circuit and truth table for A(B + C). Because two functions are equivalent if and only if they have the same output for all possible inputs, it is easy to see from the truth tables of Fig. 3-5 that A(B + C) is equivalent to AB + AC. Despite this equivalence, the circuit of Fig. 3-5(b) is clearly better than that of Fig. 3-5(a) because it contains fewer gates. In general, a circuit designer starts with a Boolean function and then applies the laws of Boolean algebra to it in an attempt to find a simpler but equivalent one. From the final function, a circuit can be constructed. To use this approach, we need some identities from Boolean algebra. Figure 3-6 shows some of the major ones. It is interesting to note that each law has two

SEC. 3.1

143

GATES AND BOOLEAN ALGEBRA AB

A B

AB + AC

A

A(B + C)

B AC

C

B+C

C

A

B

C

AB

AC

AB + AC

A

B

C

A

B+C

A(B + C)

0

0

0

0

0

0

0

0

0

0

0

0

0

0

1

0

0

0

0

0

1

0

1

0

0

1

0

0

0

0

0

1

0

0

1

0

0

1

1

0

0

0

0

1

1

0

1

0

1

0

0

0

0

0

1

0

0

1

0

0

1

0

1

0

1

1

1

0

1

1

1

1

1

1

0

1

0

1

1

1

0

1

1

1

1

1

1

1

1

1

1

1

1

1

1

1

(a)

(b)

Figure 3-5. Two equivalent functions. (a) AB + AC. (b) A(B + C).

forms that are duals of each other. By interchanging AND and OR and also 0 and 1, either form can be produced from the other one. All the laws can be easily proven by constructing their truth tables. Except for DeMorgan’s law, the absorption law, and the AND form of the distributive law, the results are reasonably intuitive. DeMorgan’s """" " " law" can be extended to more than two variables, for example, ABC = A + B + C. DeMorgan’s law suggests an alternative notation. In Fig. 3-7(a) the AND form is shown with negation indicated by inversion bubbles, both for input and output. Thus an OR gate with inverted inputs is equivalent to a NAND gate. From Fig. 3-7(b), the dual form of DeMorgan’s law, it should be clear that a NOR gate can be drawn as an AND gate with inverted inputs. By negating both forms of DeMorgan’s law, we arrive at Fig. 3-7(c) and (d), which show equivalent representations of the AND and OR gates. Analogous symbols exist for the multiple variable forms of DeMorgan’s law (e.g., an n input NAND gate becomes an OR gate with n inverted inputs). Using the identities of Fig. 3-7 and the analogous ones for multi-input gates, it is easy to convert the sum-of-products representation of a truth table to pure NAND or pure NOR form. As an example, consider the EXCLUSIVE OR function of Fig. 3-8(a). The standard sum-of-products circuit is shown in Fig. 3-8(b). To

144

THE DIGITAL LOGIC LEVEL Name

AND form

CHAP. 3 OR form

Identity law

1A = A

0+A=A

Null law

0A = 0

1+A=1

Idempotent law

AA = A

A+A=A

Inverse law

AA = 0

A+A=1

Commutative law

AB = BA

A+B=B+A

Associative law

(AB)C = A(BC)

(A + B) + C = A + (B + C)

Distributive law

A + BC = (A + B)(A + C)

A(B + C) = AB + AC

Absorption law

A(A + B) = A

A + AB = A

De Morgan's law

AB = A + B

A + B = AB

Figure 3-6. Some identities of Boolean algebra.

AB

=

A+B

A+B

(a)

AB

=

(c)

=

AB

(b)

A+B

A+B

=

AB

(d)

Figure 3-7. Alternative symbols for some gates: (a) NAND. (b) NOR. (c) AND. (d) OR.

convert to NAND form, the lines connecting the output of the AND gates to the input of the OR gate should be redrawn with two inversion bubbles, as shown in " Fig. 3-8(c). Finally, using Fig. 3-7(a), we arrive at Fig. 3-8(d). The variables A " and B can be generated from A and B using NAND or NOR gates with their inputs tied together. Note that inversion bubbles can be moved along a line at will, for example, from the outputs of the input gates in Fig. 3-8(d) to the inputs of the output gate. As a final note on circuit equivalence, we will now demonstrate the surprising result that the same physical gate can compute different functions, depending on the conventions used. In Fig. 3-9(a) we show the output of a certain gate, F, for

SEC. 3.1

145

GATES AND BOOLEAN ALGEBRA A

A

B

XOR

0

0

0

0

1

1

1

0

1

A

1

1

0

B

B

(b)

(a)

A

A

B

B

A

A

B

B (c)

(d)

Figure 3-8. (a) The truth table for the computing it.

XOR

function. (b)-(d) Three circuits for

different input combinations. Both inputs and outputs are shown in volts. If we adopt the convention that 0 volts is logical 0 and 3.3 volts or 5 volts is logical 1, called positive logic, we get the truth table of Fig. 3-9(b), the AND function. If, however, we adopt negative logic, which has 0 volts as logical 1 and 3.3 volts or 5 volts as logical 0, we get the truth table of Fig. 3-9(c), the OR function. A

B

F

A

B

F

A

B

F

0V

0V

0V

0

0

0

1

1

1

V

V

V

0

1

0

1

0

1

0

5

0

5V

0V

0V

1

0

0

0

1

1

5V

5V

5V

1

1

1

0

0

0

(a)

(b)

(c)

Figure 3-9. (a) Electrical characteristics of a device. (b) Positive logic. (c) Negative logic.

Thus the convention chosen to map voltages onto logical values is critical. Except where otherwise specified, we will henceforth use positive logic, so the terms logical 1, true, and high are synonyms, as are logical 0, false, and low.

146

THE DIGITAL LOGIC LEVEL

CHAP. 3

3.2 BASIC DIGITAL LOGIC CIRCUITS In the previous sections we saw how to implement truth tables and other simple circuits using individual gates. In practice, few circuits are actually constructed gate-by-gate anymore, although this once was common. Nowadays, the usual building blocks are modules containing a number of gates. In the following sections we will examine these building blocks more closely and see how they are used and how they can be constructed from individual gates.

3.2.1 Integrated Circuits Gates are not manufactured or sold individually but rather in units called Integrated Circuits, often called ICs or chips. An IC is a square piece of silicon about 5 mm × 5 mm on which some gates have been deposited. Small ICs are usually mounted in rectangular plastic or ceramic packages measuring 5 to 15 mm wide and 20 to 50 mm long. Along the long edges are two parallel rows of pins about 5 mm long that can be inserted into sockets or soldered to printed circuit boards. Each pin connects to the input or output of some gate on the chip or to power or to ground. The packages with two rows of pins outside and ICs inside are technically known as Dual Inline Packages or DIPs, but everyone calls them chips, thus blurring the distinction between the piece of silicon and its package. The most common packages have 14, 16, 18, 20, 22, 24, 28, 40, 64, or 68 pins. For large chips, square packages with pins on all four sides or on the bottom are often used. Chips can be divided into rough classes based on the number of gates they contain, as given below. This classification scheme is obviously extremely crude, but it is sometimes useful. SSI (Small Scale Integrated) circuit: 1 to 10 gates. MSI (Medium Scale Integrated) circuit: 10 to 100 gates. LSI (Large Scale Integrated) circuit: 100 to 100,000 gates. VLSI (Very Large Scale Integrated) circuit: >100,000 gates. These classes have different properties and are used in different ways. An SSI chip typically contains two to six independent gates, each of which can be used individually, in the style of the previous sections. Figure 3-10 illustrates a schematic drawing of a common SSI chip containing four NAND gates. Each of these gates has two inputs and one output, requiring a total of 12 pins for the four gates. In addition, the chip needs power (Vcc ), and ground (GND), which are shared by all gates. The package generally has a notch near pin 1 to identify the orientation. To avoid clutter in circuit diagrams, neither power, nor ground, nor unused gates are conventionally shown. Many other chips like this are available for a few cents each. Each SSI chip has a handful of gates and up to 20 or so pins. In the 1970s, computers were con-

SEC. 3.2

147

BASIC DIGITAL LOGIC CIRCUITS VCC 14

13

12

11

10

9

8

1

2

3

4

5

6

7

Pin 8

Notch

GND

Figure 3-10. An SSI chip containing four gates.

structed out of large numbers of these chips, but nowadays an entire CPU and a substantial amount of (cache) memory is etched onto a single chip. For our purposes, all gates are ideal in the sense that the output appears as soon as the input is applied. In reality, chips have a finite gate delay, which includes both the signal propagation time through the chip and the switching time. Typical delays are 1 to 10 nsec. It is within the current state of the art to put almost 10 million transistors on a chip. Because any circuit can be built up from NAND gates, you might think that a manufacturer could make a very general chip containing 5 million NAND gates. Unfortunately, such a chip would need 15,000,002 pins. With the standard pin spacing of 0.1 inch, the chip would be over 19 km long, which might have a negative effect on sales. Clearly, the only way to take advantage of the technology is to design circuits with a high gate/pin ratio. In the following sections we will look at simple MSI circuits that combine a number of gates internally to provide a useful function requiring only a limited number of external connections (pins).

3.2.2 Combinational Circuits Many applications of digital logic require a circuit with multiple inputs and multiple outputs in which the outputs are uniquely determined by the current inputs. Such a circuit is called a combinational circuit. Not all circuits have this property. For example, a circuit containing memory elements may well generate outputs that depend on the stored values as well as the input variables. A circuit

148

THE DIGITAL LOGIC LEVEL

CHAP. 3

implementing a truth table, such as that of Fig. 3-3(a), is a typical example of a combinational circuit. In this section we will examine some frequently-used combinational circuits. Multiplexers At the digital logic level, a multiplexer is a circuit with 2n data inputs, one data output, and n control inputs that select one of the data inputs. The selected data input is ‘‘gated’’ (i.e., routed) to the output. Figure 3-11 is a schematic diagram for an eight-input multiplexer. The three control lines, A, B, and C, encode a 3-bit number that specifies which of the eight input lines is gated to the OR gate and thence to the output. No matter what value is on the control lines, seven of the AND gates will always output 0; the other one may output either 0 or 1, depending on the value of the selected input line. Each AND gate is enabled by a different combination of the control inputs. The multiplexer circuit is shown in Fig. 3-11. When power and ground are added, it can be packaged in a 14-pin package. Using the multiplexer, we can implement the majority function of Fig. 3-3(a), as shown in Fig. 3-12(b). For each combination of A, B, and C, one of the data input lines is selected. Each input is wired to either Vcc (logical 1) or ground (logical 0). The algorithm for wiring the inputs is simple: input Di is the same as the value in row i of the truth table. In Fig. 3-3(a), rows 0, 1, 2, and 4 are 0, so the corresponding inputs are grounded; the remaining rows are 1, so they are wired to logical 1. In this manner any truth table of three variables can be implemented using the chip of Fig. 3-12(a). We just saw how a multiplexer chip can be used to select one of several inputs and how it can implement a truth table. Another of its applications is as a parallel-to-serial data converter. By putting 8 bits of data on the input lines and then stepping the control lines sequentially from 000 to 111 (binary), the 8 bits are put onto the output line in series. A typical use for parallel-to-serial conversion is in a keyboard, where each keystroke implicitly defines a 7- or 8-bit number that must be output serially over a telephone line. The inverse of a multiplexer is a demultiplexer, which routes its single input signal to one of 2n outputs, depending on the values of the n control lines. If the binary value on the control lines is k, output k is selected. Decoders As a second example, we will now look at a circuit that takes an n-bit number as input and uses it to select (i.e., set to 1) exactly one of the 2n output lines. Such a circuit, illustrated for n = 3 in Fig. 3-13, is called a decoder. To see where a decoder might be useful, imagine a small memory consisting of eight chips, each containing 1 MB. Chip 0 has addresses 0 to 1 MB, chip 1 has

SEC. 3.2

149

BASIC DIGITAL LOGIC CIRCUITS D0 D1 D2 D3 F

D4 D5 D6 D7 A A B B C C

A

B

C

Figure 3-11. An eight-input multiplexer circuit.

addresses 1 MB to 2 MB, and so on. When an address is presented to the memory, the high-order 3 bits are used to select one of the eight chips. Using the circuit of Fig. 3-13, these 3 bits are the three inputs, A, B, and C. Depending on the inputs, exactly one of the eight output lines, D 0 , ..., D 7 , is 1; the rest are 0. Each output line enables one of the eight memory chips. Because only one output line is set to 1, only one chip is enabled. The operation of the circuit of Fig. 3-13 is"straightforward. Each AND" gate has three inputs, of which"the first is either A or A, the second is either B or B, and the third is either " " C " or C. Each " " gate is enabled by a different combination of inputs: D 0 by A B C, D 1 by A B C, and so on. Comparators Another useful circuit is the comparator, which compares two input words. The simple comparator of Fig. 3-14 takes two inputs, A and B, each of length 4 bits, and produces a 1 if they are equal and a 0 if they are not equal. The circuit is

150

THE DIGITAL LOGIC LEVEL

CHAP. 3

VCC

D0

D0

D1

D1

D2

D2

D3

F

D4

D3

D5

D5

D6

D6

D7

D7

A B C (a)

F

D4

A B C (b)

Figure 3-12. (a) An MSI multiplexer.. (b) The same multiplexer wired to compute the majority function.

based on the XOR (EXCLUSIVE OR) gate, which puts out a 0 if its inputs are equal and a 1 if they are unequal. If the two input words are equal, all four of the XOR gates must output 0. These four signals can then be ORed together; if the result is 0, the input words are equal, otherwise not. In our example we have used a NOR gate as the final stage to reverse the sense of the test: 1 means equal, 0 means unequal. Programmable Logic Arrays We saw earlier that arbitrary functions (truth tables) can be constructed by computing product terms with AND gates and then ORing the products together. A very general chip for forming sums of products is the Programmable Logic Array or PLA, a small example of which is shown in Fig. 3-15. This chip has input lines for 12 variables. The complement of each input is generated internally, making 24 input signals in all. The heart of the circuit is an array of 50 AND gates, each of which can potentially have any subset of the 24 input signals as an input. Which input signal goes to which AND gate is determined by a 24 × 50 bit matrix supplied by the user. Each input line to the 50 AND gates contains a fuse. When shipped from the factory, all 1200 fuses are intact. To program the matrix the user burns out selected fuses by applying a high voltage to the chip. The output part of the circuit consists of six OR gates, each of which has up to 50 inputs, corresponding to the 50 outputs of the AND gates. Again here, a user-

SEC. 3.2

151

BASIC DIGITAL LOGIC CIRCUITS

D0

D1

A

B

A

D2

A

D3

B

D4

B C

C C

D5

D6

D7

Figure 3-13. A 3-to-8 decoder circuit.

supplied (50 × 6) matrix tells which of the potential connections actually exist. The chip has 12 input pins, 6 output pins, power, and ground, for a total of 20. As an example of how a PLA can be used, let us reconsider the circuit of Fig. 3-3(b) again. It has three inputs, four AND gates, one OR gate, and three inverters. With the appropriate internal connections made, our PLA can compute the same function using three of its 12 inputs, four of its " 50 AND " gates, " and one of its six OR gates. (The four AND gates should compute ABC, ABC, ABC, and ABC, respectively; the OR gate takes these four product terms as input.) In fact, the same PLA could be wired up to compute simultaneously a total of four functions of similar complexity. For these simple functions the number of input variables is the limiting factor; for more complicated ones it might be the AND or OR gates. Although the field-programmable PLAs described above are still in use, for many applications custom-made PLAs are preferable. These are designed by the (large-volume) customer and fabricated by the manufacturer to the customer’s specifications. Such PLAs are cheaper than field-programmable ones. We can now compare the three different ways we have discussed for implementing the truth table of Fig. 3-3(a). Using SSI components, we need four chips. Alternatively, we could suffice with one MSI multiplexer chip, as shown in Fig. 3-12(b). Finally, we could use a quarter of one PLA chip. Obviously, if

152

THE DIGITAL LOGIC LEVEL

CHAP. 3

EXCLUSIVE OR gate A0 B0

A1 B1 A=B A2 B2

A3 B3

Figure 3-14. A simple 4-bit comparator.

many functions are needed, the PLA is more efficient than the other two methods. For simple circuits, the cheaper SSI and MSI chips may be preferable.

3.2.3 Arithmetic Circuits It is now time to move on from the general-purpose MSI circuits discussed above to MSI combinational circuits used for doing arithmetic. We will begin with a simple 8-bit shifter, then look at how adders are constructed, and finally examine arithmetic logic units, which play a central role in any computer. Shifters Our first arithmetic MSI circuit is an eight-input, eight-output shifter (see Fig. 3-16). Eight bits of input are presented on lines D 0 , ..., D 7 . The output, which is just the input shifted 1 bit, is available on lines S 0 , ..., S 7 . The control line, C, determines the direction of the shift, 0 for left and 1 for right. On a left shift, a 0 is inserted into bit 7. Similarly, on a right shift, a 1 is inserted into bit 0. To see how the circuit works, notice the pairs of AND gates for all the bits except the gates on the end. When C = 1, the right member of each pair is turned on, passing the corresponding input bit to output. Because the right AND gate is wired to the input of the OR gate to its right, a right shift is performed. When C = 0, it is the left member of the AND gate pair that turns on, doing a left shift.

SEC. 3.2

153

BASIC DIGITAL LOGIC CIRCUITS

A

If this fuse is blown, B is not an input to AND gate 1

B 12 2 = 24 input signals

L

24 input lines 0

1

49

0

1 If this fuse is blown, AND gate 1 is not an input to OR gate 5

50 input lines

6 outputs

5

Figure 3-15. A 12-input, 6-output programmable logic array. The little squares represent fuses that can be burned out to determine the function to be computed. The fuses are arranged in two matrices: the upper one for the AND gates and the lower one for the OR gates.

Adders A computer that cannot add integers is almost unthinkable. Consequently, a hardware circuit for performing addition is an essential part of every CPU. The truth table for addition on 1-bit integers is shown in Fig. 3-17(a). Two outputs are present: the sum of the inputs, A and B, and the carry to the next (leftward) position. A circuit for computing both the sum bit and the carry bit is illustrated in Fig. 3-17(b). This simple circuit is generally known as a half adder.

154

THE DIGITAL LOGIC LEVEL

CHAP. 3

D0

D1

D2

D3

D4

D5

D6

D7

S0

S1

S2

S3

S4

S5

S6

S7

C

Figure 3-16. A 1-bit left/right shifter.

Exclusive OR gate A

B

0

0

0

0

0

1

1

0

1

0

1

0

1

1

0

1

Sum Carry

A

Sum

B

Carry

Figure 3-17. (a) Truth table for 1-bit addition. (b) A circuit for a half adder.

Although a half adder is adequate for summing the low-order bits of two multibit input words, it will not do for a bit position in the middle of the word because it does not handle the carry into the position from the right. Instead, the full adder of Fig. 3-18 is needed. From inspection of the circuit it should be clear that a full adder is built up from two half adders. The Sum output line is 1 if an odd number of A, B, and the Carry in are 1. The Carry out is 1 if either A and B are both 1 (left input to the OR gate) or exactly one of them is 1 and the Carry in bit is also 1. Together the two half adders generate both the sum and the carry bits.

SEC. 3.2

155

BASIC DIGITAL LOGIC CIRCUITS Carry in Carry Carry Sum in out

A

B

0

0

0

0

0

0

0

1

1

0

0

1

0

1

0

0

1

1

0

1

1

0

0

1

0

1

0

1

0

1

1

1

0

0

1

1

1

1

1

1

A

Sum

B

Carry out (a)

(b)

Figure 3-18. (a) Truth table for full adder. (b) Circuit for a full adder.

To build an adder for, say, two 16-bit words, one just replicates the circuit of Fig. 3-18(b) 16 times. The carry out of a bit is used as the carry into its left neighbor. The carry into the rightmost bit is wired to 0. This type of adder is called a ripple carry adder, because in the worst case, adding 1 to 111...111 (binary), the addition cannot complete until the carry has rippled all the way from the rightmost bit to the leftmost bit. Adders that do not have this delay, and hence are faster, also exist and are usually preferred. As a simple example of a faster adder, consider breaking a 32-bit adder up into a 16-bit lower half and a 16-bit upper half. When the addition starts, the upper adder cannot yet get to work because it will not know the carry into it for 16 addition times. However, consider this modification. Instead of having a single upper half, give the adder two upper halves in parallel by duplicating the upper half’s hardware. Thus the circuit now consists of three 16-bit adders: a lower half and two upper halves, U0 and U1 that run in parallel. A 0 is fed into U0 as a carry; a 1 is fed into U1 as a carry. Now both of these can start at the same time the lower half starts, but only one will be correct. After 16 bit-addition times, it will be known what the carry into the upper half is, so the correct upper half can now be selected from the two available answers. This trick reduces the addition time by a factor of two. Such an adder is called a carry select adder. This trick can then be repeated to build each 16-bit adder out of replicated 8-bit adders, and so on.

156

THE DIGITAL LOGIC LEVEL

CHAP. 3

Arithmetic Logic Units Most computers contain a single circuit for performing the AND, OR, and sum of two machine words. Typically, such a circuit for n-bit words is built up of n identical circuits for the individual bit positions. Figure 3-19 is a simple example of such a circuit, called an Arithmetic Logic Unit or"ALU. It can compute any one of four functions—namely, A AND B, A OR B, B, or A + B, depending on whether the function-select input lines F 0 and F 1 contain 00, 01, 10, or 11 (binary). Note that here A + B means the arithmetic sum of A and B, not the Boolean OR. Logical unit

Carry in

AB INVA A ENA B ENB

A+B

Output

B Sum

Enable lines

F0

Full adder

F1

Decoder

Carry out

Figure 3-19. A 1-bit ALU.

The lower left-hand corner of our ALU contains a 2-bit decoder to generate enable signals for the four operations, based on the control signals F 0 and F 1 . Depending on the values of F 0 and F 1 exactly one of the four enable lines is

SEC. 3.2

157

BASIC DIGITAL LOGIC CIRCUITS

selected. Setting this line allows the output for the selected function to pass through to the final OR gate for output. " The upper left-hand corner has the logic to compute A AND B, A OR B, and B, but at most one of these results is passed onto the final OR gate, depending on the enable lines coming out of the decoder. Because exactly one of the decoder outputs will be 1, exactly one of the four AND gates driving the OR gate will be enabled; the other three will output 0, independent of A and B. In addition to being able to use A and B as inputs for logical or arithmetic operations, it is also possible to force "either one to 0 by negating ENA or ENB, respectively. It is also possible to get A, by setting INVA. We will see uses for INVA, ENA , and ENB in Chap. 4. Under normal conditions, ENA and ENB are both 1 to enable both inputs and INVA is 0. In this case, A and B are just fed into the logic unit unmodified. The lower right-hand corner of the ALU contains a full adder for computing the sum of A and B, including handling the carries, because it is likely that several of these circuits will eventually be wired together to perform full-word operations. Circuits like Fig. 3-19 are actually available and are known as bit slices. They allow the computer designer to build an ALU of any desired width. Figure 3-20 shows an 8-bit ALU built up of eight 1-bit ALU slices. The INC signal is only useful for addition operations. When present, it increments (i.e., adds 1 to) the result, making it possible to compute sums like A + 1 and A + B + 1. F1 F 0

A7 B7

A 6 B6

A 5 B5

A4 B4

A3 B3

A 2 B2

A 1 B1

A0 B0

1-bit ALU

1-bit ALU

1-bit ALU

1-bit ALU

1-bit ALU

1-bit ALU

1-bit ALU

1-bit ALU

O7

O6

O5

O4

O3

O2

O1

O0

Carry in

INC

Carry out

Figure 3-20. Eight 1-bit ALU slices connected to make an 8-bit ALU. The enables and invert signals are not shown for simplicity.

3.2.4 Clocks In many digital circuits the order in which events happen is critical. Sometimes one event must precede another, sometimes two events must occur simultaneously. To allow designers to achieve the required timing relations, many digital circuits use clocks to provide synchronization. A clock in this context is a circuit that emits a series of pulses with a precise pulse width and precise interval between consecutive pulses. The time interval between the corresponding edges of two consecutive pulses is known as the clock cycle time. Pulse frequencies are

158

THE DIGITAL LOGIC LEVEL

CHAP. 3

commonly between 1 and 500 MHz, corresponding to clock cycles of 1000 nsec to 2 nsec. To achieve high accuracy, the clock frequency is usually controlled by a crystal oscillator. In a computer, many events may happen during a single clock cycle. If these events must occur in a specific order, the clock cycle must be divided into subcycles. A common way of providing finer resolution than the basic clock is to tap the primary clock line and insert a circuit with a known delay in it, thus generating a secondary clock signal that is phase-shifted from the primary, as shown in Fig. 3-21(a). The timing diagram of Fig. 3-21(b) provides four time references for discrete events: 1. Rising edge of C1. 2. Falling edge of C1. 3. Rising edge of C2. 4. Falling edge of C2. By tying different events to the various edges, the required sequencing can be achieved. If more than four time references are needed within a clock cycle, more secondary lines can be tapped from the primary, with different delays. C1 Delay

C2

(a)

(b)

A B C (c)

Figure 3-21. (a) A clock. (b) The timing diagram for the clock. (c) Generation of an asymmetric clock.

In some circuits, one is interested in time intervals rather than discrete instants of time. For example, some event may be allowed to happen any time C1 is high, rather than precisely at the rising edge. Another event may only happen when C2

SEC. 3.2

159

BASIC DIGITAL LOGIC CIRCUITS

is high. If more than two different intervals are needed, more clock lines can be provided or the high states of the two clocks can be made to overlap"partially " "in " AND C2 , time. In the latter case four distinct intervals can be distinguished: C1 "" "" C1 AND C2, C1 AND C2, and C1 AND C2. As an aside, clocks are symmetric, with time spent in the high state equal to the time spent in the low state, as shown in Fig. 3-21(b). To generate an asymmetric pulse train, the basic clock is shifted using a delay circuit and ANDed with the original signal, as shown in Fig. 3-21(c) as C.

3.3 MEMORY An essential component of every computer is its memory. Without memory there could be no computers as we now know them. Memory is used for storing both instructions to be executed and data. In the following sections we will examine the basic components of a memory system starting at the gate level to see how they work and how they are combined to produce large memories.

3.3.1 Latches To create a 1-bit memory, we need a circuit that somehow ‘‘remembers’’ previous input values. Such a circuit can be constructed from two NOR gates, as illustrated in Fig. 3-22(a). Analogous circuits can be built from NAND gates. We will not mention these further, however, because they are conceptually identical to the NOR versions. S

0

1

Q

S

0

0

Q

1

1 R

0

0 0

0 (a)

Q

R

1

0 (b)

Q

A

B

NOR

0

0

1

0

1

0

1

0

0

1

1

0

(c)

Figure 3-22. (a) NOR latch in state 0. (b) NOR latch in state 1. (c) Truth table for NOR.

The circuit of Fig. 3-22(a) is called an SR latch. It has two inputs, S, for Setting "the " latch, and R, for Resetting (i.e., clearing) it. It also has two outputs, Q and Q, which are complementary, as we will see shortly. Unlike a combinational circuit, the outputs of the latch are not uniquely determined by the current inputs.

160

THE DIGITAL LOGIC LEVEL

CHAP. 3

To see how this comes about, let us assume that both S and R are 0, which they are most of the time. For argument’s sake, let us further assume that Q = 0. Because "" Q is fed back into the upper NOR gate, both of its inputs are 0, so its output, Q, is 1. The 1 is fed back into the lower gate, which then has inputs 1 and 0, yielding Q = 0. This state is at least consistent and is depicted in Fig. 3-22(a). Now let us imagine that Q is not 0"" but 1, with R and S still 0. The upper gate has inputs of 0 and 1, and an output, Q, of 0, which is fed back to the lower gate. This state, shown in Fig. 3-22(b), is also consistent. A state with both outputs equal to 0 is inconsistent, because it forces both gates to have two 0s as input, which, if true, would produce 1, not 0, as output. Similarly, it is impossible to have both outputs equal to 1, because that would force the inputs to 0 and 1, which yields 0, not 1. Our conclusion is simple: for R = S = 0, the latch has two stable states, which we will refer to as 0 and 1, depending on Q. Now let us examine the effect of the inputs on the state of the latch. Suppose 1 while Q = 0. The inputs to the upper gate are then 1 and 0, forcthat S becomes "" ing the Q output to 0. This change makes both inputs to the lower gate 0, forcing the output to 1. Thus setting S (i.e., making it 1) switches the state from 0 to 1. Setting R to 1 when the latch is in state 0 has no effect because the output of the lower NOR gate is 0 for inputs of 10 and inputs of 11. Using similar reasoning, it is easy to see that setting S to 1 when in state Q = 1 has no effect but that setting R drives the latch to state Q = 0. In summary, when S is set to 1 momentarily, the latch ends up in state Q = 1, regardless of what state it was previously in. Likewise, setting R to 1 momentarily forces the latch to state Q = 0. The circuit ‘‘remembers’’ whether S or R was last on. Using this property we can build computer memories. Clocked SR Latches It is often convenient to prevent the latch from changing state except at certain specified times. To achieve this goal, we modify the basic circuit slightly, as shown in Fig. 3-23, to get a clocked SR latch. S Q Clock Q R

Figure 3-23. A clocked SR latch.

This circuit has an additional input, the clock, which is normally 0. With the clock 0, both AND gates output 0, independent of S and R, and the latch does not change state. When the clock is 1, the effect of the AND gates vanishes and the

SEC. 3.3

161

MEMORY

latch becomes sensitive to S and R. Despite its name, the clock signal need not be driven by a clock. The terms enable and strobe are also widely used to mean that the clock input is 1; that is, the circuit is sensitive to the state of S and R. Up until now we have carefully swept the problem of what happens when both S and R are 1 under the rug. And for good reason: the circuit becomes nondeterministic when "" both R and S finally return to 0. The only consistent state for S = R = 1 is Q = Q = 0, but as soon as both inputs return to 0, the latch must jump to one of its two stable states. If either input drops back to 0 before the other, the one remaining 1 longest wins, because when just one input is 1, it forces the state. If both inputs return to 0 simultaneously (which is very unlikely), the latch jumps to one of its stable states at random. Clocked D Latches A good way to resolve the SR latch’s ambiguity (caused when S = R = 1) is to prevent it from occurring. Figure 3-24 gives a latch circuit with only one input, D. Because the input to the lower AND gate is always the complement of the input to the upper one, the problem of both inputs being 1 never arises. When D = 1 and the clock is 1, the latch is driven into state Q = 1. When D = 0 and the clock is 1, it is driven into state Q = 0. In other words, when the clock is 1, the current value of D is sampled and stored in the latch. This circuit, called a clocked D latch, is a true 1-bit memory. The value stored is always available at Q. To load the current value of D into the memory, a positive pulse is put on the clock line. D Q

Q

Figure 3-24. A clocked D latch.

This circuit requires 11 transistors. More sophisticated (but less obvious) circuits can store 1 bit with as few a six transistors. In practice, such designs are normally used.

3.3.2 Flip-Flops In many circuits it is necessary to sample the value on a certain line at a particular instant in time and store it. In this variant, called a flip-flop, the state transition does not occur when the clock is 1 but during the clock transition from 0 to

162

THE DIGITAL LOGIC LEVEL

CHAP. 3

1 (rising edge) or from 1 to 0 (falling edge) instead. Thus the length of the clock pulse is unimportant, as long as the transitions occur fast. For emphasis, we will repeat the difference between a flip-flop and a latch. A flip-flop is edge triggered, whereas a latch is level triggered. Be warned, however, that in the literature these terms are often confused. Many authors use ‘‘flip-flop’’ when they are referring to a latch, and vice versa. There are various approaches to designing a flip-flop. For example, if there were some way to generate a very short pulse on the rising edge of the clock signal, that pulse could be fed into a D latch. There is, in fact, such a way, and the circuit for it is shown in Fig. 3-25(a).

d ∆ a

b c

(a)

d

b AND c c b a Time (b)

Figure 3-25. (a) A pulse generator. (b) Timing at four points in the circuit.

At first glance, it might appear that the output of the AND gate would always be zero, since the AND of any signal with its inverse is zero, but the situation is a bit more subtle than that. The inverter has a small, but nonzero propagation delay through it, and that delay is what makes the circuit work. Suppose that we measure the voltage at the four measuring points a, b, c, and d. The input signal, measured at a, is a long clock pulse, as shown in Fig. 3-25(b) on the bottom. The signal at b is shown above it. Notice that it is both inverted and delayed slightly, typically a few nanoseconds depending on the kind of inverter used. The signal at c is delayed, too, but only by the signal propagation time (at the speed of light). If the physical distance between a and c is, for example, 20 microns, then the propagation delay is 0.0001 nsec, which is certainly negligible compared to the time for the signal to propagate through the inverter. Thus for all intents and purposes, the signal at c is as good as identical to the signal at a.

SEC. 3.3

MEMORY

163

When the inputs to the AND gate, b and c, are ANDed together, the result is a short pulse, as shown in Fig. 3-25(b), where the width of the pulse, ∆, is equal to the gate delay of the inverter, typically 5 nsec or less. The output of the AND gate is just this pulse shifted by the delay of the AND gate, as shown at the top of Fig. 3-25(b). This time shifting just means that the D latch will be activated at a fixed delay after the rising edge of the clock, but it has no effect on the pulse width. In a memory with a 50-nsec cycle time, a 5-nsec pulse telling it when to sample the D line may be short enough, in which case the full circuit can be the one of Fig. 3-26. It is worth noting that this flip-flop design is nice because it is easy to understand, but in practice more sophisticated flip-flops are normally used. D Q

Q

Figure 3-26. A D flip-flop.

The standard symbols for latches and flip-flops are shown in Fig. 3-27. Figure 3-27(a) is a latch whose state is loaded when the clock, CK, is 1, in contrast to Fig. 3-27(b) which is a latch whose clock is normally 1 but which drops to 0 momentarily to load the state from D. Figure 3-27(c) and (d) are flip-flops rather than latches, which is indicated by the pointy symbol on the clock inputs. Figure 3-27(c) changes state on the rising edge of the clock pulse (0 to 1 transition), whereas Fig. 3-27(d) changes state on the falling "" edge (1 to 0 transition). Many, but not all, latches and flip-flops also have Q as an output, and some have two additional inputs Set or Preset (force state to Q = 1) and Reset or Clear (force state to Q = 0).

3.3.3 Registers Flip-flops are available in a variety of configurations. A simple one, containing two independent D flip-flops with clear and preset signals, is illustrated in Fig. 3-28(a). Although packaged together in the same 14-pin chip, the two flipflops are unrelated. A quite different arrangement is the octal flip-flop of Fig. 328(b). "" Here the eight (hence the term ‘‘octal’’) D flip-flops are not only missing the Q and preset lines, but all the clock lines are ganged together and driven by

164

THE DIGITAL LOGIC LEVEL

D

Q

CK (a)

D

Q

CK (b)

D

Q

CK (c)

CHAP. 3

D

Q

CK (d)

Figure 3-27. D latches and flip-flops.

pin 11. The flip-flops themselves are of the Fig. 3-27(d) type, but the inversion bubbles on the flip-flops are canceled by the inverter tied to pin 11, so the flipflops are loaded on the rising transition. All eight clear signals are also ganged, so when pin 1 goes to 0, all the flip-flops are forced to their 0 state. In case you are wondering why pin 11 is inverted at the input and then inverted again at each CK signal, an input signal may not have enough current to drive all eight flip-flops; the input inverter is really being used as an amplifier. While one reason for ganging the clock and clear lines of Fig. 3-28(b) is to save pins, in this configuration the chip is used in a different way from eight unrelated flip-flops. It is used as a single 8-bit register. Alternatively, two such chips can be used in parallel to form a 16-bit register by tying their respective pins 1 and 11 together. We will look at registers and their uses more closely in Chap. 4.

3.3.4 Memory Organization Although we have now progressed from the simple 1-bit memory of Fig. 3-24 to the 8-bit memory of Fig. 3-28(b) to build large memories a different organization is required, one in which individual words can be addressed. A widely-used memory organization that meets this criterion is shown in Fig. 3-29. This example illustrates a memory with four 3-bit words. Each operation reads or writes a full 3-bit word. While the total memory capacity of 12 bits is hardly more than our octal flip-flop, it requires fewer pins and most important, the design extends easily to large memories. While the memory of Fig. 3-29 may look complicated at first, it is really quite simple due to its regular structure. It has eight input lines and three output lines. Three inputs are data: I0 , I1 , and I2 ; two are for the address: A0 and A1 ; and three are for control: CS for Chip Select, RD for distinguishing between read and write, and OE for Output Enable. The three outputs are for data: O0 , O1 , and O2 . In principle this memory could be put into a 14-pin package, including power and ground versus 20 pins for the octal flip-flop. To select this memory chip, external logic must set CS high and also set RD high (logical 1) for read and low (logical 0) for write. The two address lines must be set to indicate which of the four 3-bit words is to be read or written. For a read

SEC. 3.3

165

MEMORY VCC 13

14

12

D

CLR

11

10

D

Q

2

8

Q

CK Q PR

CK Q PR

1

CLR

9

3

4

5

6

7 GND

(a) VCC 20

19

Q

2

D

17

Q

16

15

D

Q

14

D

13

Q

12

CK CLR

CK CLR

CK CLR

CLR CK

CLR CK

CLR CK

CLR CK

D

3

Q

4

D

Q

5

6

D

7

Q

8

11

D

CK CLR

Q

1

18

D

9

10 GND

(b)

Figure 3-28. (a) Dual D flip-flop. (b) Octal flip-flop.

166

THE DIGITAL LOGIC LEVEL

CHAP. 3

Data in I2 I1 I0 Write gate

Word 0 select line

A1 A0

Word 1 select line

Word 2 select line

D Q

D Q

D Q

CK

CK

CK

D Q

D Q

D Q

CK

CK

CK

D Q

D Q

D Q

CK

CK

CK

D Q

D Q

D Q

CK

CK

CK

Word 0

Word 1

Word 2

Word 3

CS • RD CS O2

RD

O1 Oo

OE

Output enable = CS • RD • OE

Figure 3-29. Logic diagram for a 4 × 3 memory. Each row is one of the four 3-bit words. A read or write operation always reads or writes a complete word.

operation, the data input lines are not used, but the word selected is placed on the data output lines. For a write operation, the bits present on the data input lines are loaded into the selected memory word; the data output lines are not used.

SEC. 3.3

MEMORY

167

Now let us look at Fig. 3-29 closely to see how it works. The four wordselect AND gates at the left of the memory form a decoder. The input inverters have been placed so that each gate is enabled (output is high) by a different address. Each gate drives a word select line, from top to bottom, for words 0, 1, 2, and """ 3. When the chip has been selected for a write, the vertical line labeled CS . RD will be high, enabling one of the four write gates, depending on which word select line is high. The output of the write gate drives all the CK signals for the selected word, loading the input data into the flip-flops for that word. A write is only done if CS is high and RD is low, and even then only the word selected by A0 and A1 is written; the other words are not changed. Read is similar to write. The address decoding is exactly the same as for """ write. But now the CS . RD line is low, so all the write gates are disabled and none of the flip-flops is modified. Instead, the word select line that is chosen enables the AND gates tied to the Q bits of the selected word. Thus the selected word outputs its data into the four-input OR gates at the bottom of the figure, while the other three words output 0s. Consequently, the output of the OR gates is identical to the value stored in the word selected. The three words not selected make no contribution to the output. Although we could have designed a circuit in which the three OR gates were just fed into the three output data lines, doing so sometimes causes problems. In particular, we have shown the data input lines and the data output lines as being different, but in actual memories the same lines are used. If we had tied the OR gates to the data output lines, the chip would try to output data, that is, force each line to a specific value, even on writes, thus interfering with the input data. For this reason, it is desirable to have a way to connect the OR gates to the data output lines on reads but disconnect them completely on writes. What we need is an electronic switch that can make or break a connection in a few nanoseconds. Fortunately, such switches exist. Figure 3-30(a) shows the symbol for what is called a noninverting buffer. It has a data input, a data output, and a control input. When the control input is high, the buffer acts like a wire, as shown in Fig. 3-30(b). When the control input is low, the buffer acts like an open circuit, as shown in Fig. 3-30(c); it is as though someone detached the data output from the rest of the circuit with a wirecutter. However, in contrast to the wirecutter approach, the connection can be subsequently restored in a few nanoseconds by just making the control signal high again. Figure 3-30(d) shows an inverting buffer, which acts like a normal inverter when control is high and disconnects the output from the circuit when control is low. Both kinds of buffers are tri-state devices, because they can output 0, 1, or none of the above (open circuit). Buffers also amplify signals, so they can drive many inputs simultaneously. They are sometimes used in circuits for this reason, even when their switching properties are not needed. Getting back to the memory circuit, it should now be clear what the three noninverting buffers on the data output lines are for. When CS, RD, and OE are all

168 Data in

THE DIGITAL LOGIC LEVEL

CHAP. 3

Data out

Control (a)

(b)

(c)

(d)

Figure 3-30. (a) A noninverting buffer. (b) Effect of (a) when control is high. (c) Effect of (a) when control is low. (d) An inverting buffer.

high, the output enable signal is also high, enabling the buffers and putting a word onto the output lines. When any one of CS, RD, or OE is low, the data outputs are disconnected from the rest of the circuit.

3.3.5 Memory Chips The nice thing about the memory of Fig. 3-29 is that it extends easily to larger sizes. As we drew it, the memory is 4 × 3, that is, four words of 3 bits each. To extend it to 4 × 8 we need only add five more columns of four flip-flops each, as well as five more input lines and five more output lines. To go from 4 × 3 to 8 × 3 we must add four more rows of three flip-flops each, as well as an address line A 2 . With this kind of structure, the number of words in the memory should be a power of 2 for maximum efficiency, but the number of bits in a word can be anything. Because integrated circuit technology is well suited to making chips whose internal structure is a repetitive two-dimensional pattern, memory chips are an ideal application for it. As the technology improves, the number of bits that can be put on a chip keeps increasing, typically by a factor of two every 18 months (Moore’s law). The larger chips do not always render the smaller ones obsolete due to different trade-offs in capacity, speed, power, price, and interfacing convenience. Commonly, the largest chips currently available sell at a premium and thus are more expensive per bit than older, smaller ones. For any given memory size, there are various ways of organizing the chip. Figure 3-31 shows two possible organizations for an older memory chip of size 4-Mbit: 512K × 8 and 4096K × 1. (As an aside, memory chip sizes are usually quoted in bits, rather than in bytes, so we will stick to that convention here.) In Fig. 3-31(a), 19 address lines are needed to address one of the 219 bytes, and eight data lines are needed for loading or storing the byte selected. A note on terminology is in order here. On some pins, the high voltage causes an action to happen. On others, the low voltage causes the action. To avoid confusion, we will consistently say that a signal is asserted (rather than saying it goes high or goes low) to mean that it is set to cause some action. Thus for some pins,

SEC. 3.3 A0 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12 A13 A14 A15 A16 A17 A18

169

MEMORY

512K 3 8 Memory chip (4 Mbit)

D0 D1 D2 D3 D4 D5 D6 D7

A0 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10

4096K 3 1 Memory chip (4 Mbit)

D

RAS CAS

CS WE OE

CS WE OE

(a)

(b)

Figure 3-31. Two ways of organizing a 4-Mbit memory chip.

asserting it means setting it high. For others, it means setting the pin low. Pins that are asserted low are given signal names "containing an overbar. Thus a signal "" named CS is asserted high, but one named CS is asserted low. The opposite of asserted is negated. When nothing special is happening, pins are negated. Now let us get back to our memory chip. Since a computer normally has many memory chips, a signal is needed to select the """chip that is currently needed so that it responds and all the others do not. The CS (Chip Select) signal is provided for this purpose. It is asserted""" to enable the chip. Also, a way is needed to is used to indicate distinguish reads from writes. The WE signal (Write Enable) """ that data are being written rather than being read. Finally, the OE (Output Enable) signal is asserted to drive the output signals. When it is not asserted, the chip output is disconnected from the circuit. In Fig. 3-31(b), a different addressing scheme is used. Internally, this chip is organized as a 2048 × 2048 matrix of 1-bit cells, which gives 4 Mbits. To address the chip, first """" a row is selected by putting its 11-bit number on the address pins. Strobe) is asserted. After that, a column number is Then the RAS (Row Address"""" put on the address pins and CAS (Column Address Strobe) is asserted. The chip responds by accepting or outputting one data bit. Large memory chips are often constructed as n × n matrices that are addressed by row and column. This organization reduces the number of pins required but also makes addressing the chip slower, since two addressing cycles are needed, one for the row and one for the column. To win back some of the speed lost by this design, some memory chips can be given a row address followed by a sequence of column addresses to access consecutive bits in a row.

170

THE DIGITAL LOGIC LEVEL

CHAP. 3

Years ago, the largest memory chips were often organized like Fig. 3-31(b). As memory words have grown from 8 bits to 32 bits and beyond, 1-bit wide chips began to be inconvenient. To build a memory with a 32-bit word from 4096K × 1 chips requires 32 chips in parallel. These 32 chips have a total capacity of at least 16 MB, whereas using 512K × 8 chips requires only four chips in parallel and allows memories as small as 2 MB. To avoid having 32 chips for memory, most chip manufacturers now have chip families with 4-, 8-, and 16-bit widths. And the situation with 64-bit words is even worse, of course. Two examples of modern 512-Mbit chips are given in Fig. 3-32. These chips have four internal memory banks of 128 Mbit each, requiring two bank select lines to choose """"a bank. The design of Fig. """"3-32(a) is a 32M × 16 design, with 13 lines for the RAS signal, 10 lines for the CAS signal, and 2 lines for the bank select. adTogether, these 25 signals allow each of the 225 internal 16-bit cells to be """" RAS a 128M the is for 3-32(b) lines 13 with dressed. In contrast, Fig. × 4 design, """" signal, 12 lines for the CAS signal, and 2 lines for the bank select. Here, 27 signals can select any of the 227 internal 4-bit cells to be addressed. The decision about how many rows and how many columns a chip has is made for engineering reasons. The matrix need not be square. A0 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12

32M 3 16 Memory chip (512 Mbit)

RAS CAS Bank 0 Bank 1

D0 D1 D2 D3 D4 D5 D6 D7 D8 D9 D10 D11 D12 D13 D14 D15

A0 A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12

128M 3 4 Memory chip (512 Mbit)

D0 D1 D2 D3

RAS CAS Bank 0 Bank 1

CS WE OE

CS WE OE

(a)

(b)

Figure 3-32. Two ways of organizing a 512-Mbit memory chip.

These examples demonstrate two separate and independent issues for memory chip design. First is the output width (in bits): does the chip deliver 1, 4, 8, 16, or some other number of bits at once? Second, are all the address bits presented on

SEC. 3.3

MEMORY

171

separate pins at once or are the row and columns presented sequentially as in the examples of Fig. 3-32? A memory chip designer has to answer both questions before starting the chip design.

3.3.6 RAMs and ROMs The memories we have studied so far can all be read and written. Such memories are called RAMs (Random Access Memories), which is a misnomer because all memory chips are randomly accessible, but the term is too well established to get rid of now. RAMs come in two varieties, static and dynamic. Static RAMs (SRAMs), are constructed internally using circuits similar to our basic D flip-flop. These memories have the property that their contents are retained as long as the power is kept on: seconds, minutes, hours, even days. Static RAMs are very fast. A typical access times is a few nsec. For this reason, static RAMS are popular as level 2 cache memory. Dynamic RAMs (DRAMs), in contrast, do not use flip-flops. Instead, a dynamic RAM is an array of cells, each cell containing one transistor and a tiny capacitor. The capacitors can be charged or discharged, allowing 0s and 1s to be stored. Because the electric charge tends to leak out, each bit in a dynamic RAM must be refreshed (reloaded) every few milliseconds to prevent the data from leaking away. Because external logic must take care of the refreshing, dynamic RAMs require more complex interfacing than static ones, although in many applications this disadvantage is compensated for by their larger capacities. Since dynamic RAMs need only one transistor and one capacitor per bit (vs. six transistors per bit for the best static RAM), dynamic RAMs have a very high density (many bits per chip). For this reason, main memories are nearly always built out of dynamic RAMs. However, this large capacity has a price: dynamic RAMs are slow (tens of nanoseconds). Thus the combination of a static RAM cache and a dynamic RAM main memory attempts to combine the good properties of each. Several types of dynamic RAM chips exist. The oldest type still around (in elderly computers) is FPM (Fast Page Mode) DRAM. Internally it is organized as a matrix of bits and it works by having the hardware present"""" a row address """" and then step through the column addresses, as we described with RAS and CAS in the context of Fig. 3-31. Explicit signals tell the memory when it is time to respond, so the memory runs asynchronously from the main system clock. FPM DRAM was replaced with EDO (Extended Data Output) DRAM, which allows a second memory reference to begin before the previous memory reference has been completed. This simple pipelining did not make a single memory reference go faster but did improve the memory bandwidth, giving more words per second. FPM and EDO worked reasonably well when memory chips had cycle times of 12 nsec and slower. When processors got so fast that faster memories were

172

THE DIGITAL LOGIC LEVEL

CHAP. 3

really needed, FPM and EDO were replaced by SDRAM (Synchronous DRAM), which is a hybrid of static and dynamic RAM and is driven by the main system clock. The big advantage of SDRAM is that the clock eliminates the need for control signals to tell the memory chip when to respond. Instead, the CPU tells the memory how many cycles it should run, then starts it. On each subsequent cycle, the memory outputs 4, 8, or 16 bits, depending on how many output lines it has. Eliminating the need for control signals increases the data rate between CPU and memory. The next improvement over SDRAM was DDR (Double Data Rate) SDRAM. With this kind of memory, the memory chip produces output on both the rising edge of the clock and the falling edge, doubling the data rate. Thus an 8-bit wide DDR chip running at 200 MHz outputs two 8-bit values 200 million times a second (for a short interval, of course), giving a theoretical burst rate of 3.2 Gbps. Nonvolatile Memory Chips RAMs are not the only kind of memory chips. In many applications, such as toys, appliances, and cars, the program and some of the data must remain stored even when the power is turned off. Furthermore, once installed, neither the program nor the data are ever changed. These requirements have led to the development of ROMs (Read-Only Memories), which cannot be changed or erased, intentionally or otherwise. The data in a ROM are inserted during its manufacture, essentially by exposing a photosensitive material through a mask containing the desired bit pattern and then etching away the exposed (or unexposed) surface. The only way to change the program in a ROM is to replace the entire chip. ROMs are much cheaper than RAMs when ordered in large enough volumes to defray the cost of making the mask. However, they are inflexible, because they cannot be changed after manufacture, and the turnaround time between placing an order and receiving the ROMs may be weeks. To make it easier for companies to develop new ROM-based products, the PROM (Programmable ROM) was invented. A PROM is like a ROM, except that it can be programmed (once) in the field, eliminating the turnaround time. Many PROMs contain an array of tiny fuses inside. A specific fuse can be blown out by selecting its row and column and then applying a high voltage to a special pin on the chip. The next development in this line was the EPROM (Erasable PROM), which can be not only field-programmed but also field-erased. When the quartz window in an EPROM is exposed to a strong ultraviolet light for 15 minutes, all the bits are set to 1. If many changes are expected during the design cycle, EPROMs are far more economical than PROMs because they can be reused. EPROMs usually have the same organization as static RAMs. The 4-Mbit 27C040 EPROM, for example, uses the organization of Fig. 3-32(a), which is typical of a static RAM.

SEC. 3.3

MEMORY

173

Even better than the EPROM is the EEPROM which can be erased by applying pulses to it instead of requiring it to be put in a special chamber for exposure to ultraviolet light. In addition, an EEPROM can be reprogrammed in place whereas an EPROM has to be inserted in a special EPROM programming device to be programmed. On the minus side, the biggest EEPROMs are typically only 1/64 as large as common EPROMs and they are only half as fast. EEPROMs cannot compete with DRAMs or SRAMs because they are 10 times slower, 100 times smaller in capacity, and much more expensive. They are only used in situations where their nonvolatility is crucial. A more recent kind of EEPROM is flash memory. Unlike EPROM, which is erased by exposure to ultraviolet light, and EEPROM, which is byte erasable, flash memory is block erasable and rewritable. Like EEPROM, flash memory can be erased without removing it from the circuit. Various manufacturers produce small printed circuit cards with up to 1 GB of flash memory on them for use as ‘‘film’’ for storing pictures in digital cameras and many other purposes. Someday flash memories may be used to replace disks, which would be an enormous improvement, given their 50-nsec access times. The main engineering problem at present is that they wear out after about 100,000 erasures, whereas disks last for many years, no matter how often they are rewritten. A summary of the various kinds of memory is given in Fig. 3-33. #################################################################################### ! ! ! ! Byte ! ! ! ! ! ! ! ! ! ! Erasure Typical use #################################################################################### ! Type ! Category ! ! alterable ! Volatile ! ! !#################################################################################### ! Read/write ! Electrical ! Yes ! Yes ! Level 2 cache ! SRAM ! ! ! ! ! ! ! ! DRAM ! Read/write ! Electrical ! Yes ! Yes ! Main memory (old) ! #################################################################################### ! SDRAM ! Read/write ! Electrical ! Yes ! Yes ! Main memory (new) ! #################################################################################### ! ! ! ! ! ! ! ! ROM ! Read-only ! Not possible ! No ! No ! Large volume appliances ! #################################################################################### ! PROM ! Read-only ! Not possible ! No ! No ! Small volume equipment ! #################################################################################### ! ! ! ! ! ! ! !#################################################################################### ! No ! No ! Device prototyping ! EPROM ! Read-mostly ! UV light ! ! ! ! ! ! ! EEPROM ! Read-mostly ! Electrical !#################################################################################### ! Yes ! No ! Device prototyping ! !#################################################################################### !! Read/write !! Electrical !! No !! No !! Film for digital camera !! ! Flash Figure 3-33. A comparison of various memory types.

3.4 CPU CHIPS AND BUSES Armed with information about SSI chips, MSI chips, and memory chips, we can now start to put all the pieces together to look at complete systems. In this section, we will first look at some general aspects of CPUs as viewed from the digital logic level, including pinout (what the signals on the various pins mean).

174

THE DIGITAL LOGIC LEVEL

CHAP. 3

Since CPUs are so closely intertwined with the design of the buses they use, we will also provide an introduction to bus design in this section. In succeeding sections we will give detailed examples of both CPUs and their buses and how they are interfaced.

3.4.1 CPU Chips All modern CPUs are contained on a single chip. This makes their interaction with the rest of the system well defined. Each CPU chip has a set of pins, through which all its communication with the outside world must take place. Some pins output signals from the CPU to the outside world; others accept signals from the outside world; some can do both. By understanding the function of all the pins, we can learn how the CPU interacts with the memory and I/O devices at the digital logic level. The pins on a CPU chip can be divided into three types: address, data, and control. These pins are connected to similar pins on the memory and I/O chips via a collection of parallel wires called a bus. To fetch an instruction, the CPU first puts the memory address of that instruction on its address pins. Then it asserts one or more control lines to inform the memory that it wants to read (for example) a word. The memory replies by putting the requested word on the CPU’s data pins and asserting a signal saying that it is done. When the CPU sees this signal, it accepts the word and carries out the instruction. The instruction may require reading or writing data words, in which case the whole process is repeated for each additional word. We will go into the detail of how reading and writing works below. For the time being, the important thing to understand is that the CPU communicates with the memory and I/O devices by presenting signals on its pins and accepting signals on its pins. No other communication is possible. Two of the key parameters that determine the performance of a CPU are the number of address pins and the number of data pins. A chip with m address pins can address up to 2m memory locations. Common values of m are 16, 20, 32 and 64. Similarly, a chip with n data pins can read or write an n-bit word in a single operation. Common values of n are 8, 16, 32, 36, and 64. A CPU with 8 data pins will take four operations to read a 32-bit word, whereas one with 32 data pins can do the same job in one operation. Thus the chip with 32 data pins is much faster, but is invariably more expensive as well. In addition to address and data pins, each CPU has some control pins. The control pins regulate the flow and timing of data to and from the CPU and have other miscellaneous uses. All CPUs have pins for power (usually +3.3 volts or +5 volts), ground, and a clock signal (a square wave at some well-defined frequency), but the other pins vary greatly from chip to chip. Nevertheless, the control pins can be roughly grouped into the following major categories:

SEC. 3.4

CPU CHIPS AND BUSES

175

1. Bus control. 2. Interrupts. 3. Bus arbitration. 4. Coprocessor signaling. 5. Status. 6. Miscellaneous. We will briefly describe each of these categories below. When we look at the Pentium 4, UltraSPARC III, and 8051 chips later, we will provide more detail. A generic CPU chip using these signal groups is shown in Fig. 3-34. Addressing Data Bus control

Bus arbitration Coprocessor

Typical MicroProcessor

Status

Interrupts

Symbol for clock signal

Miscellaneous

Φ +5v

Symbol for electrical ground

Power is 5volts

Figure 3-34. The logical pinout of a generic CPU. The arrows indicate input signals and output signals. The short diagonal lines indicate that multiple pins are used. For a specific CPU, a number will be given to tell how many.

The bus control pins are mostly outputs from the CPU to the bus (thus inputs to the memory and I/O chips) telling whether the CPU wants to read or write memory or do something else. The CPU uses these pins to control the rest of the system and tell it what it wants to do. The interrupt pins are inputs from I/O devices to the CPU. In most systems, the CPU can tell an I/O device to start an operation and then go off and do something else while the I/O device is doing its work. When the I/O has been completed, the I/O controller chip asserts a signal on one of these pins to interrupt the CPU and have it service the I/O device, for example to check to see if I/O errors occurred. Some CPUs have an output pin to acknowledge the interrupt signal. The bus arbitration pins are needed to regulate traffic on the bus, in order to prevent two devices from trying to use it at the same time. For arbitration purposes, the CPU counts as a device and has to request it like any other device.

176

THE DIGITAL LOGIC LEVEL

CHAP. 3

Some CPU chips are designed to operate with coprocessors such as floatingpoint chips, but sometimes graphics or other chips as well. To facilitate communication between CPU and coprocessor, special pins are provided for making and granting various requests. In addition to these signals, there are various miscellaneous pins that some CPUs have. Some of these provide or accept status information, others are useful for resetting the computer, and still others are present to assure compatibility with older I/O chips.

3.4.2 Computer Buses A bus is a common electrical pathway between multiple devices. Buses can be categorized by their function. They can be used internal to the CPU to transport data to and from the ALU, or external to the CPU, to connect it to memory or to I/O devices. Each type of bus has its own requirements and properties. In this section and the following ones, we will focus on buses that connect the CPU to the memory and I/O devices. In the next chapter we will examine the buses inside the CPU more closely. Early personal computers had a single external bus or system bus. It consisted of 50 to 100 parallel copper wires etched onto the motherboard, with connectors spaced at regular intervals for plugging in memory and I/O boards. Modern personal computers generally have a special-purpose bus between the CPU and memory and (at least) one other bus for the I/O devices. A minimal system, with one memory bus and one I/O bus, is illustrated in Fig. 3-35. CPU chip Buses Registers

Memory bus

Bus controller

I/O bus

ALU

On-chip bus

Memory

Disk

Modem

Printer

Figure 3-35. A computer system with multiple buses.

In the literature, buses are sometimes drawn as ‘‘fat’’ arrows, as in this figure. The distinction between a fat arrow and a single line with a diagonal line through it and a bit count next to it is subtle. When all the bits are the same type, say, all

SEC. 3.4

CPU CHIPS AND BUSES

177

address bits or all data bits, then the short diagonal line approach is commonly used. When there are address, data, and control lines involved, a fat arrow is more common. While the designers of the CPU are free to use any kind of bus they want inside the chip, in order to make it possible for boards designed by third parties to attach to the system bus, there must be well-defined rules about how the bus works, and which all devices attached to it must obey. These rules are called the bus protocol. In addition, there must be mechanical and electrical specifications, so that third-party boards will fit in the card cage and have connectors that match those on the motherboard mechanically and in terms of voltages, timing, etc. A number of buses are in widespread use in the computer world. A few of the better known ones, historical and current, (with examples) are the Omnibus (PDP-8), Unibus (PDP-11), Multibus (8086), VME bus (physics lab equipment), IBM PC bus (PC/XT), ISA bus (PC/AT), EISA bus (80386), Microchannel (PS/2), Nubus (Macintosh), PCI bus (many PCs), SCSI bus (many PCs and workstations), Universal Serial Bus (modern PCs), and FireWire (consumer electronics). The world would probably be a better place if all but one would suddenly vanish from the face of the earth (well, all right, how about all but two?). Unfortunately, standardization in this area seems very unlikely as there is already too much invested in all these incompatible systems. Let us now begin our study of how buses work. Some devices that attach to a bus are active and can initiate bus transfers, whereas others are passive and wait for requests. The active ones are called masters; the passive ones are called slaves. When the CPU orders a disk controller to read or write a block, the CPU is acting as a master and the disk controller is acting as a slave. However, later on, the disk controller may act as a master when it commands the memory to accept the words it is reading from the disk drive. Several typical combinations of master and slave are listed in Fig. 3-36. Under no circumstances can memory ever be a master. ############################################################################## ! ! ! Master Slave Example !############################################################################## ! ! ! CPU !############################################################################## ! Memory ! Fetching instructions and data ! ! CPU ! I/O device ! Initiating data transfer ! !############################################################################## ! ! ! !############################################################################## ! Coprocessor ! CPU handing instruction off to coprocessor ! CPU ! ! ! ! I/O ! Memory ! DMA (Direct Memory Access) ! !############################################################################## ! ! !! !############################################################################## Coprocessor CPU Coprocessor fetching operands from CPU ! ! ! !

Figure 3-36. Examples of bus masters and slaves.

The binary signals that computer devices output are frequently too weak to power a bus, especially if it is relatively long or has many devices on it. For this reason, most bus masters are connected to the bus by a chip called a bus driver, which is essentially a digital amplifier. Similarly, most slaves are connected to

178

THE DIGITAL LOGIC LEVEL

CHAP. 3

the bus by a bus receiver. For devices that can act as both master and slave, a combined chip called a bus transceiver is used. These bus interface chips are often tri-state devices, to allow them to float (disconnect) when they are not needed, or are hooked up in a somewhat different way, called open collector, that achieves a similar effect. When two or more devices on an open collector line assert the line at the same time, the result is the Boolean OR of all the signals. This arrangement is often called wired-OR. On most buses, some of the lines are tri-state and others, which need the wired-OR property, are open collector. Like a CPU, a bus also has address, data, and control lines. However, there is not necessarily a one-to-one mapping between the CPU pins and the bus signals. For example, some CPUs have three pins that encode whether it is doing a memory read, memory write, I/O read, I/O write, or some other operation. A typical bus might have one line for memory read, a second for memory write, a third for I/O read, a fourth for I/O write, and so on. A decoder chip would then be needed between the CPU and such a bus to match the two sides up, that is, to convert the 3-bit encoded signal into separate signals that can drive the bus lines. Bus design and operation are sufficiently complex subjects that a number of entire books have been written about them (Anderson et al., 2004; Solari and Willse, 2004). The principal bus design issues are bus width, bus clocking, bus arbitration, and bus operations. Each of these issues has a substantial impact on the speed and bandwidth of the bus. We will now examine each of these in the next four sections.

3.4.3 Bus Width Bus width is the most obvious design parameter. The more address lines a bus has, the more memory the CPU can address directly. If a bus has n address lines, then a CPU can use it to address 2n different memory locations. To allow large memories, buses need many address lines. That sounds simple enough. The problem is that wide buses need more wires than narrow ones. They also take up more physical space (e.g., on the motherboard) and need bigger connectors. All of these factors make the bus more expensive. Thus there is a trade-off between maximum memory size and system cost. A system with a 64-line address bus and 232 bytes of memory will cost more than one with 32 address lines and the same 232 bytes of memory. The possibility of expansion later is not free. The result of this observation is that many system designers tend to be shortsighted, with unfortunate consequences later. The original IBM PC contained an 8088 CPU and a 20-bit address bus, as shown in Fig. 3-37(a). Having 20 bits allowed the PC to address 1 MB of memory. When the next CPU chip (the 80286) came out, Intel decided to increase the address space to 16 MB, so four more bus lines had to be added (without disturbing the original 20, for reasons of backward compatibility), as illustrated in Fig. 3-37(b). Unfortunately, more control lines had to be added to deal with the

SEC. 3.4

179

CPU CHIPS AND BUSES

20-Bit address 20-Bit address Control

20-Bit address 8088

Control 80286

4-Bit address 80386 Control 8-Bit address

4-Bit address

Control

Control Control (a)

(b)

(c)

Figure 3-37. Growth of an address bus over time.

new address lines. When the 80386 came out, another eight address lines were added, along with still more control lines, as shown in Fig. 3-37(c). The resulting design (the EISA bus) is much messier than it would have been had the bus been given 32 lines at the start. Not only does the number of address lines tend to grow over time, but so does the number of data lines, but for a somewhat different reason. There are two ways to increase the data bandwidth of a bus: decrease the bus cycle time (more transfers/sec) or increase the data bus width (more bits/transfer). Speeding the bus up is possible, but difficult because the signals on different lines travel at slightly different speeds, a problem known as bus skew. The faster the bus, the more serious bus skew becomes. Another problem with speeding up the bus is that doing this will not be backward compatible. Old boards designed for the slower bus will not work with the new one. Invalidating old boards makes both the owners of the old boards and the manufacturers of the old boards unhappy. Therefore the usual approach to improving performance is to add more data lines, analogous to Fig. 3-37. As you might expect, however, this incremental growth does not lead to a clean design in the end. The IBM PC and its successors, for example, went from eight data lines to 16 and then 32 on essentially the same bus. To get around the problem of very wide buses, sometimes designers opt for a multiplexed bus. In this design, instead of having the address and data lines be separate, there are, say, 32 lines for address and data together. At the start of a bus operation, the lines are used for the address. Later on, they are used for data. For a write to memory, for example, this means that the address lines must be set up and propagated to the memory before the data can be put on the bus. With separate lines, the address and data can be put on together. Multiplexing the lines reduces bus width (and cost), but results in a slower system. Bus designers have to carefully weigh all these options when making choices.

180

THE DIGITAL LOGIC LEVEL

CHAP. 3

3.4.4 Bus Clocking Buses can be divided into two distinct categories depending on their clocking. A synchronous bus has a line driven by a crystal oscillator. The signal on this line consists of a square wave with a frequency generally between 5 MHz and 100 MHz. All bus activities take an integral number of these cycles, called bus cycles. The other kind of bus, the asynchronous bus, does not have a master clock. Bus cycles can be of any length required and need not be the same between all pairs of devices. Below we will examine each bus type. Synchronous Buses As an example of how a synchronous bus works, consider the timing of Fig. 3-38(a). In this example, we will use a 100-MHz clock, which gives a bus cycle of 10 nsec. While this may seem a bit slow compared to CPU speeds of 3 GHz and more, few existing PC buses are much faster. For example, the popular PCI bus usually runs at either 33 MHz or 66 MHz. The reasons current buses are slow were given above: technical design problems such as bus skew and the need for backward compatibility. In our example, we will further assume that reading from memory takes 15nsec from the time the address is stable. As we will see shortly, with these parameters, it will take three bus cycles to read a word. The first cycle starts at the rising edge of T1 and the third one ends at the rising edge of T4 , as shown in the figure. Note that none of the rising or falling edges has been drawn vertically, because no electrical signal can change its value in zero time. In this example we will assume """""" that """ it takes " """"1 nsec for a signal to change. The clock, ADDRESS, DATA , MREQ, RD, and WAIT lines are all shown on the same time scale. The start of T1 is defined by the rising edge of the clock. Part way through T1 the CPU puts the address of the word it wants on the address lines. Because the address is not a single value, like the clock, we cannot show it as a single line in the figure; instead, it is shown as two lines, with a crossing at the time that the address changes. Furthermore, the shading prior to the crossing indicates that the shaded value is not important. Using the same shading convention, we see that the contents of the data lines are not significant until well into T3 . After the """""" """ address lines have had a chance to settle down to their new values, MREQ and RD are asserted. The former indicates that memory (as opposed to an I/O device) is being accessed, and the latter is asserted for reads and negated for writes. Since the memory takes 15 nsec after the address is stable (part way into the first clock cycle), it cannot provide the requested " """" data during T2 . To tell the CPU not to expect it, the memory asserts the WAIT line at the start of T2 . This insert wait states (extra bus cycles) until the memory is finished and action will " """" negates WAIT . In our example, one wait state (T2 ) has been inserted because the when it is sure it will have the data during memory is too slow. At the start of T3", """" the current cycle, the memory negates WAIT .

SEC. 3.4

181

CPU CHIPS AND BUSES Read cycle with 1 wait state T1

Φ

T2

T3

TAD

ADDRESS

Memory address to be read

TDS DATA

Data TM

MREQ

TMH

TML TRH

RD

TDH

TRL WAIT Time (a)

Symbol

Parameter

Min

TAD

Address output delay

TML

Address stable prior to MREQ

TM

MREQ delay from falling edge of Φ in T1

Max

Unit

4

nsec

2

nsec 3

nsec

3

nsec

TRL

RD delay from falling edge of Φ in T1

TDS

Data setup time prior to falling edge of Φ

TMH

MREQ delay from falling edge of Φ in T3

3

nsec

TRH

RD delay from falling edge of Φ in T3

3

nsec

TDH

Data hold time from negation of RD

2

nsec

0

nsec

(b)

Figure 3-38. (a) Read timing on a synchronous bus. (b) Specification of some critical times.

During the first half of T3 , the memory puts the data onto the data lines. At the falling edge of T3 the CPU strobes (i.e., reads) the data lines, latching (i.e., storing) value in an internal register. Having read the data, the CPU negates """""" the""" MREQ and RD. If need be, another memory cycle can begin at the next rising edge of the clock. This sequence can be repeatedly indefinitely.

182

THE DIGITAL LOGIC LEVEL

CHAP. 3

In the timing specification of Fig. 3-38(b), eight symbols that occur in the timing diagram are further clarified. TAD , for example, is the time interval between the rising edge of the T1 clock and the address lines being set. According to the timing specification, TAD ≤ 4 nsec. This means that the CPU manufacturer guarantees that during any read cycle, the CPU will output the address to be read within 11 nsec of the midpoint of the rising edge of T1 . The timing specifications also require that the data be available on the data lines at least TDS (2 nsec) before the falling edge of T3 , to give it time to settle down before the CPU strobes it in. The combination of constraints on TAD and TDS means that, in the worst case, the memory will have only 25 − 4 − 2 = 19 nsec from the time the address appears until it must produce the data. Because 10 nsec is enough, even in the worst case, a 10-nsec memory can always respond during T3 . A 20-nsec memory, however, would just miss and have to insert a second wait state and respond during T4 . The timing specification up at """""" further guarantees that the address will be set """""" least 2 nsec prior to MREQ being asserted. This time can be important if MREQ drives chip select on the memory chip because some memories require an address setup time prior to chip select. Clearly, the system designer should not choose a memory chip that needs a 3-nsec setup time. """""" """ The constraints on TM and TRL mean that MREQ and RD will both be asserted the memory within 3 nsec from the T1 falling clock. In the worst case,"""""" """ chip will have only 10 + 10 − 3 − 2 = 15 nsec after the assertion of MREQ and RD to get its data onto the bus. This constraint is in addition to (and independent of) the 15nsec interval needed after the address is stable. """""" """ TMH and TRH tell how long it takes MREQ and RD to be negated after the data have been strobed """in. Finally, TDH tells how long the memory must hold the data on the bus after RD has been negated. As far as our example """ CPU is concerned, the memory can remove the data from the bus as soon as RD has been negated; on some actual CPUs, however, the data must be kept stable a little longer. We would like to point out that Fig. 3-38 is a highly simplified version of real timing constraints. In reality, many more critical times are always specified. Nevertheless, it gives a good flavor for how a synchronous bus works. A last point worth making is that control signals can be asserted high or low. It is up to the bus designers to determine which is more convenient, but the choice is essentially arbitrary. One can regard it as the hardware equivalent of a programmer’s choice to represent free disk blocks in a bit map as 0s versus 1s. Asynchronous Buses Although synchronous buses are easy to work with due to their discrete time intervals, they also have some problems. For example, everything works in multiples of the bus clock. If a CPU and memory are able to complete a transfer in 3.1 cycles, they have to stretch it to 4.0 because fractional cycles are forbidden.

SEC. 3.4

183

CPU CHIPS AND BUSES

Worse yet, once a bus cycle has been chosen, and memory and I/O cards have been built for it, it is difficult to take advantage of future improvements in technology. For example, suppose a few years after the system of Fig. 3-38 was built, new memories became available with access times of 8 nsec instead of 15 nsec. These would get rid of the wait state, speeding up the machine. Then suppose 4nsec memories became available. There would be no further gain in performance because the minimum time for a read is two cycles with this design. Putting this fact in slightly different terms, if a synchronous bus has a heterogeneous collection of devices, some fast and some slow, the bus has to be geared to the slowest one and the fast ones cannot use their full potential. Mixed technology can be handled by going to an asynchronous bus, that is, one with no master clock, as shown in Fig. 3-39. Instead"""""" of tying """ everything to and anything the clock, when the bus master has asserted the address, MREQ, RD,"""""" else it needs to, it then asserts a special signal that we will call MSYN (Master SYNchronization). When the slave """"" sees this, it performs the work as fast as it can. When it is done, it asserts SSYN (Slave SYNchronization). ADDRESS

Memory address to be read

MREQ RD MSYN

Data

DATA SSYN

Figure 3-39. Operation of an asynchronous bus. """"" master sees SSYN asserted, it knows that the

As soon as the data"""""" are available, """ MREQ , RD, and so it latches them, and then negates the address lines, along with """""" """""" MSYN. When the slave sees """"" the negation of MSYN, it knows that the cycle has been completed, so it negates SSYN, and we are back in the original situation, with all signals negated, waiting for the next master. Timing diagrams of asynchronous buses (and sometimes synchronous buses of well) use arrows to show cause and effect, as in Fig. 3-39. The assertion as """""" """"" data lines to be asserted and also causes the slave to assert MSYN causes the """"" SSYN """""" """ . SSYN , in turn, causes the negation of the address lines, MREQ , RD , The assertion of """""" """""" """"" and MSYN. Finally, the negation of MSYN causes the negation of SSYN, which ends the read and returns the system to its original state.

184

THE DIGITAL LOGIC LEVEL

CHAP. 3

A set of signals that interlocks this way is called a full handshake. The essential part consists of four events: """"""

1. MSYN is asserted. """""

""""""

""""""

"""""

2. SSYN is asserted in response to MSYN. 3. MSYN is negated in response to SSYN. """""

""""""

4. SSYN is negated in response to the negation of MSYN. It should be clear that full handshakes are timing independent. Each event is caused by a prior event, not by a clock pulse. If a particular master-slave pair is slow, that in no way affects a subsequent master-slave pair that is much faster. The advantage of an asynchronous bus should now be clear, but the fact is that most buses are synchronous. The reason is that it is easier to build a synchronous system. The CPU just asserts its signals, and the memory just reacts. There is no feedback (cause and effect), but if the components have been chosen properly, everything will work without handshaking. Also, there is a lot of investment in synchronous bus technology.

3.4.5 Bus Arbitration Up until now, we have tacitly assumed that there is only one bus master, the CPU. In reality, I/O chips have to become bus master to read and write memory, and also to cause interrupts. Coprocessors may also need to become bus master. The question then arises: ‘‘What happens if two or more devices all want to become bus master at the same time?’’ The answer is that some bus arbitration mechanism is needed to prevent chaos. Arbitration mechanisms can be centralized or decentralized. Let us first consider centralized arbitration. One particularly simple form of centralized arbitration is shown in Fig. 3-40(a). In this scheme, a single bus arbiter determines who goes next. Many CPUs have the arbiter built into the CPU chip, but sometimes a separate chip is needed. The bus contains a single wired-OR request line that can be asserted by one or more devices at any time. There is no way for the arbiter to tell how many devices have requested the bus. The only categories it can distinguish are some requests and no requests. When the arbiter sees a bus request, it issues a grant by asserting the bus grant line. This line is wired through all the I/O devices in series, like a cheap string of Christmas tree lamps. When the device physically closest to the arbiter sees the grant, it checks to see if it made a request. If so, it takes over the bus but does not propagate the grant further down the line. If it has not made a request, it propagates the grant to the next device in line, which behaves the same way, and so on until some device accepts the grant and takes the bus. This scheme is called

SEC. 3.4

185

CPU CHIPS AND BUSES Bus request

Arbiter

Bus grant

Bus grant may or may not be propagated along the chain

1

2

3

4

5

3

4

5

I/O devices (a)

Arbiter

Bus request level 1 Bus request level 2 Bus grant level 2 Bus grant level 1

1

2 (b)

Figure 3-40. (a) A centralized one-level bus arbiter using daisy chaining. (b) The same arbiter, but with two levels.

daisy chaining. It has the property that devices are effectively assigned priorities depending on how close to the arbiter they are. The closest device wins. To get around the implicit priorities based on distance from the arbiter, many buses have multiple priority levels. For each priority level there is a bus request line and a bus grant line. The one of Fig. 3-40(b) has two levels, 1 and 2 (real buses often have 4, 8, or 16 levels). Each device attaches to one of the bus request levels, with more time-critical devices attaching to the higher priority ones. In Fig. 3-40(b) devices, 1, 2, and 4 use priority 1 while devices 3 and 5 use priority 2. If multiple priority levels are requested at the same time, the arbiter issues a grant only on the highest priority one. Among devices of the same priority, daisy chaining is used. In Fig. 3-40(b), in the event of conflicts, device 2 beats device 4, which beats 3. Device 5 has the lowest priority because it is at the end of the lowest priority daisy chain. As an aside, it is not technically necessary to wire the level 2 bus grant line serially through devices 1 and 2, since they cannot make requests on it. However, as an implementation convenience, it is easier to wire all the grant lines through all the devices, rather than making special wiring that depends on which device has which priority. Some arbiters have a third line that a device asserts when it has accepted a grant and seized the bus. As soon as it has asserted this acknowledgement line,

186

THE DIGITAL LOGIC LEVEL

CHAP. 3

the request and grant lines can be negated. As a result, other devices can request the bus while the first device is using the bus. By the time the current transfer is finished, the next bus master will have already been selected. It can start as soon as the acknowledgement line has been negated, at which time the following round of arbitration can begin. This scheme requires an extra bus line and more logic in each device, but it makes better use of bus cycles. In systems in which memory is on the main bus, the CPU must compete with all the I/O devices for the bus on nearly every cycle. One common solution for this situation is to give the CPU the lowest priority, so it gets the bus only when nobody else wants it. The idea here is that the CPU can always wait, but I/O devices frequently must acquire the bus quickly or lose incoming data. Disks rotating at high speed cannot wait. This problem is avoided in many modern computer systems by putting the memory on a separate bus from the I/O devices so they do not have to compete for access to the bus. Decentralized bus arbitration is also possible. For example, a computer could have 16 prioritized bus request lines. When a device wants to use the bus, it asserts its request line. All devices monitor all the request lines, so at the end of each bus cycle, each device knows whether it was the highest priority requester, and thus whether it is permitted to use the bus during the next cycle. Compared to centralized arbitration, this arbitration method requires more bus lines but avoids the potential cost of the arbiter. It also limits the number of devices to the number of request lines. Another kind of decentralized bus arbitration, shown in Fig. 3-41, only uses three lines, no matter how many devices are present. The first bus line is a wired-OR line for requesting the bus. The second bus line is called BUSY and is asserted by the current bus master. The third line is used to arbitrate the bus. It is daisy chained through all the devices. The head of this chain is held asserted by tying it to the 5-volt power supply. Bus request Busy +5v Arbitration line

In Out

In Out

In Out

In Out

In Out

1

2

3

4

5

Figure 3-41. Decentralized bus arbitration.

When no device wants the bus, the asserted arbitration line is propagated through to all devices. To acquire the bus, a device first checks to see if the bus is idle and the arbitration signal it is receiving, IN, is asserted. If IN is negated, it may not become bus master, and it negates OUT. If IN is asserted, however, the device negates OUT, which causes its downstream neighbor to see IN negated and to negate its OUT. Hence, downstream devices all see IN negated and correspond-

SEC. 3.4

CPU CHIPS AND BUSES

187

ingly negate OUT. When the dust settles, only one device will have IN asserted and OUT negated. This device becomes bus master, asserts BUSY and OUT, and begins its transfer. Some thought will reveal that the leftmost device that wants the bus gets it. Thus this scheme is similar to the original daisy chain arbitration, except without having the arbiter, so it is cheaper, faster, and not subject to arbiter failure.

3.4.6 Bus Operations Up until now, we have only discussed ordinary bus cycles, with a master (typically the CPU) reading from a slave (typically the memory) or writing to one. In fact, several other kinds of bus cycles exist. We will now look at some of these. Normally, one word at a time is transferred. However, when caching is used, it is desirable to fetch an entire cache line (e.g., 16 consecutive 32-bit words) at once. Often block transfers can be made more efficient than successive individual transfers. When a block read is started, the bus master tells the slave how many words are to be transferred, for example, by putting the word count on the data lines during T1 . Instead of just returning one word, the slave outputs one word during each cycle until the count has been exhausted. Figure 3-42 shows a modi""""""" fied version of Fig. 3-38(a), but now with an extra signal BLOCK that is asserted to indicate that a block transfer is requested. In this example, a block read of 4 words takes 6 cycles instead of 12. Other kinds of bus cycles also exist. For example, on a multiprocessor system with two or more CPUs on the same bus, it is often necessary to make sure that only one CPU at a time uses some critical data structure in memory. A typical way to arrange this is to have a variable in memory that is 0 when no CPU is using the data structure and 1 when it is in use. If a CPU wants to gain access to the data structure, it must read the variable, and if it is 0, set it to 1. The trouble is, with some bad luck, two CPUs might read it on consecutive bus cycles. If each one sees that the variable is 0 then each one sets it to 1 and thinks that it is the only CPU using the data structure. This sequence of events leads to chaos. To prevent this situation, multiprocessor systems often have a special readmodify-write bus cycle that allows any CPU to read a word from memory, inspect and modify it, and write it back to memory, all without releasing the bus. This type of cycle prevents competing CPUs from being able to use the bus and thus interfere with the first CPU’s operation. Another important kind of bus cycle is for handling interrupts. When the CPU commands an I/O device to do something, it usually expects an interrupt when the work is done. The interrupt signaling requires the bus. Since multiple devices may want to cause an interrupt simultaneously, the same kind of arbitration problems are present here that we had with ordinary bus cycles. The usual solution is to assign priorities to devices, and use a centralized

188

THE DIGITAL LOGIC LEVEL

T1

T2

T3

CHAP. 3

T4

T5

T6

Data

Data

Data

T7

Φ ADDRESS DATA

Memory address to be read Count

Data

MREQ RD WAIT BLOCK

Figure 3-42. A block transfer.

arbiter to give priority to the most time-critical devices. Standard interrupt controller chips exist and are widely used. The IBM PC and all its successors use the Intel 8259A chip, illustrated in Fig. 3-43. INT INTA CPU

RD WR A0 CS

8259A Interrupt controller

D0-D7

IR0 IR1 IR2 IR3 IR4 IR5 IR6 IR7

Clock Keyboard

Disk Printer

+5 v

Figure 3-43. Use of the 8259A interrupt controller.

Up to eight I/O controller chips can be directly connected to the eight IRx (Interrupt Request) inputs to the 8259A. When any of these devices wants to cause an interrupt, it asserts its input line. When one or more inputs are asserted, the 8259A asserts INT (INTerrupt), which directly drives the interrupt pin on the CPU. When the CPU is able to handle the interrupt, it sends a pulse back to the 8259A on INTA (INTerrupt Acknowledge). At that point the 8259A must specify

SEC. 3.4

CPU CHIPS AND BUSES

189

which input caused the interrupt by outputting that input’s number on the data bus. This operation requires a special bus cycle. The CPU hardware then uses that number to index into a table of pointers, called interrupt vectors, to find the address of the procedure to run to service the interrupt. The 8259A has several registers """inside of it """that the CPU """ can read and write RD (ReaD), WR (WRite), CS (Chip Select), and using ordinary bus cycles and the """ A0 pins. When the software has handled the interrupt and is ready to take the next one, it writes a special code into one of the registers, which causes the 8259A to negate INT, unless it has another interrupt pending. These registers can also be written to put the 8259A in one of several modes, mask out a set of interrupts, and enable other features. When more than eight I/O devices are present, the 8259As can be cascaded. In the most extreme case, all eight inputs can be connected to the outputs of eight more 8259As, allowing for up to 64 I/O devices in a two-stage interrupt network. The 8259A has a few pins devoted for handling this cascading, which we have omitted for the sake of simplicity. While we have by no means exhausted the subject of bus design, the material above should give enough background to understand the essentials of how a bus works, and how CPUs and buses interact. Let us now move from the general to the specific and take a look at some examples of actual CPUs and their buses.

3.5 EXAMPLE CPU CHIPS In this section we will examine the Pentium 4, the UltraSPARC III, and the 8051 chips in some detail at the hardware level.

3.5.1 The Pentium 4 The Pentium 4 is a direct descendant of the 8088 CPU used in the original IBM PC. The first Pentium 4 was introduced in Nov. 2000 as a 42-million transistor CPU running at 1.5 GHz with a line width of 0.18 micron. The line width is how wide the wires between transistors are (as well as being a measure of the size of the transistors themselves). The narrower the line width, the more transistors can fit on the chip. Moore’s law is fundamentally about the ability of process engineers to keep reducing the line widths. Smaller line widths also allow higher clock speeds. For comparison purposes, human hairs range from 20 microns to 100 microns in diameter, with blonde hair being finer than black hair. During the course of the next three years, as Intel gained experience with the manufacturing process, it evolved to having 55 million transistors running at speeds up to 3.2 GHz with line widths of 0.09 micron. Although the Pentium 4 is a far cry from the 29,000-transistor 8088, it is fully backward compatible with the

190

THE DIGITAL LOGIC LEVEL

CHAP. 3

8088 and can run unmodified 8088 binary programs (not to mention programs for all the intermediate processors as well). From a software point of view, the Pentium 4 is a full 32-bit machine. It has all the same user-level ISA features as the 80386, 80486, Pentium, Pentium II, Pentium Pro, and Pentium III, including the same registers, same instructions, and a full on-chip implementation of the IEEE 754 floating-point standard. In addition, though, it has some new instructions intended primarily for multimedia applications. However, from a hardware perspective, Pentium 4 is partially a 64-bit machine since it can transfer data to and from memory in units of 64 bits. Although the programmer cannot observe these 64-bit transfers, they do make the machine faster than a pure 32-bit machine would be. Internally, at the microarchitecture level, the Pentium 4 is radically different from all its predecessors. Its immediate predecessors— the Pentium II, the Pentium Pro, and the Pentium III—all used the same internal microarchitecture (called P6), differing only in speed and in a few minor ways. In contrast, the Pentium 4 uses a new microarchitecture (called NetBurst), which is significantly different from the P6. It has a deeper pipeline, two ALUs (each of which runs at twice the clock frequency to allow two operations per cycle), and supports hyperthreading. The latter feature provides two sets of registers and some other internal resources, allowing the Pentium 4 to switch between two programs very quickly, as though the computer contained two physical CPUs. We will examine the microarchitecture in Chap. 4. However, like its predecessors, the Pentium 4 can carry out multiple instructions at once, making it a superscalar machine. Some models of the Pentium 4 have a two-level cache and some have a three-level cache. All models have an 8-KB on-chip SRAM level-one (L1) cache. Unlike Pentium III L1 cache, which just holds raw bytes from the memory, the Pentium 4 goes further. When instructions are fetched from memory, they are converted to micro-operations for actual execution in the Pentium 4 RISC core. The L1 cache on the Pentium 4 holds up to 12,000 decoded micro-operations, eliminating the need to decode them repeatedly. The second level cache holds up to 256 KB of memory in the older models and up to 1 MB of bytes in the newer ones. Nothing is decoded; pure bytes from memory are stored in the L2 cache. It can hold a mixture of code and data. The Pentium 4 Extreme Edition also has a 2-MB level 3 cache, to raise the performance even more. Since all Pentium 4 chips have at least two levels of cache, a problem arises in a multiprocessor system when one CPU has modified a word in its cache. If another CPU tries to read that word from memory, it will get a stale value since modified cache words are not written back to memory immediately. To maintain memory consistency, each CPU in a multiprocessor system snoops on the memory bus looking for references to words it has cached. When it sees such a reference, it jumps in and supplies the required data before the memory gets a chance to do so. We will study snooping in Chap. 8.

SEC. 3.5

EXAMPLE CPU CHIPS

191

35 mm

Two primary external buses are used in Pentium 4 systems, both of them synchronous. The memory bus is used to access the main (S)DRAM; the PCI bus is used for talking to I/O devices. Sometimes a legacy (i.e., ancient) bus is attached to the PCI bus to allow old peripheral devices to be plugged in. One substantial difference between the Pentium 4 and all of its predecessors is its packaging. A problem with all modern chips is the amount of power they consume and heat they produce. The Pentium 4 consumes between 63 watts and 82 watts, depending on the frequency. Consequently, Intel is constantly searching for ways to manage the heat produced by the CPU chips. The Pentium 4 comes in a square package 35 mm on edge. It contains 478 pins on the bottom, 85 of which are for power and 180 of which are grounded to reduce noise. The pins are arranged as a 26 × 26 square, with the middle 14 × 14 missing. Two pins in one corner are also missing, to prevent the chip from being inserted incorrectly in its socket. The physical pinout is shown in Fig. 3-44.

Figure 3-44. The Pentium 4 physical pinout.

The chip is outfitted with a mounting bracket for a heat sink to distribute the heat and a fan to cool it. To get some idea of what the problem is, turn on a 60watt light bulb, let it warm up, and then put your hands around it (but do not touch it). This amount of heat must be dissipated continuously. Consequently, when a Pentium 4 has outlived its usefulness as a CPU, it can always be used as a camp stove. According to the laws of physics, anything that puts out a lot of heat must suck in a lot of energy. In a portable computer with a limited battery charge, using a lot of energy is not desirable because it drains the battery quickly. To address this issue, Intel has provided a way to put the CPU to sleep when it is idle and to put it into a deep sleep when it is likely to be that way for a while. There are five states provided, ranging from fully active to deep sleep. In the intermediate states, some functionality (such as cache snooping and interrupt handling) is enabled, but other functions are disabled. When in deep sleep state, the cache and

192

THE DIGITAL LOGIC LEVEL

CHAP. 3

register values are preserved, but the clock and all the internal units are turned off. When in deep sleep, a hardware signal is required to wake it up. It is not known whether a Pentium 4 can dream when it is in deep sleep. The Pentium 4’s Logical Pinout The 478 pins on the Pentium 4 are used for 198 signals, 85 power connections (at several different voltages), 180 grounds, and 15 spares for future use. Some of the logical signals use two or more pins (such as the memory address requested), so there are only 56 different signals. A somewhat simplified logical pinout is given in Fig. 3-45. On the left side of the figure are five major groups of memory bus signals; on the right side are various miscellaneous signals. The names given entirely in uppercase are the actual Intel signal names. The ones given in mixed case are collective names for multiple related signals. Intel uses a naming convention that is important to understand. Because all chips are designed using computers these days, there is a need to be able to represent signal names as ASCII text. Using overbars to indicate signals that are asserted """""low is too difficult, so Intel puts the # symbol after the name instead. Thus BPRI is expressed as BPRI#. As can be seen from the figure, most Pentium 4 signals are asserted low. Let us examine the signals, starting with the bus signals. The first group of signals is used to request the bus (i.e., do bus arbitration). BR0 is used to request the bus. BPRI# allows a device to make a high priority request, which takes precedence over a regular one. LOCK# allows a CPU to lock the bus, to prevent other devices from getting in until it is done. Once bus ownership has been acquired, a CPU or other bus master can make a bus request using the next group of signals. Addresses are 36 bits, but the loworder 3 bits must always be 0 and therefore do not have pins assigned, so A# has only 33 pins. All transfers are 8 bytes, aligned on an 8-byte boundary. With 36 address bits, the maximum addressable memory is 236 , which is 64 GB. When an address is put onto the bus, the ADS# signal is asserted to tell the target (e.g., the memory) that the address lines are valid. The type of bus cycle (e.g., read one word or write a block) goes on the REQ# lines. The two parity lines protect A# and REQ#. The five error lines are used to report floating-point errors, internal errors, machine check (i.e., hardware) errors, and certain other errors. The Response group contains signals used by the slave to report back to the is master. RS# contains the status code. TRDY# indicates that the slave (the target) """" ready to accept data from the master. These signals are also parity checked. BNR is used to assert a wait state when the target addressed cannot respond on time. The last bus group is for the actual data transfer. D# is used to put 8 data bytes onto the bus. When they are placed there, DRDY# is asserted to announce their presence. DBSY# is used to tell the world that the bus is currently busy. Parity is

SEC. 3.5

193

EXAMPLE CPU CHIPS

Bus arbitration

Request

Error

Response

Data

BPRI# LOCK# BR0# A# ADS# REQ# Parity# Misc RS# TRDY# Parity# BNR# D# DRDY# DBSY# Parity# Misc

RESET# 33

3

5 2

14 4

5

5

Pentium 4 CPU

2

23 2 2

64

7

4 13

85 Φ

Interrupts Power management Heat management Clock frequency Diagnostics Initialization Snooping Miscellaneous

180

Power

Figure 3-45. Logical pinout of the Pentium 4. Names in uppercase are the official Intel names for individual signals. Names in mixed case are groups of related signals or signal descriptions.

also used here to check the data. The miscellaneous data signals deal with latching values and similar things. RESET# is used to reset the CPU in the event of a calamity or when the user presses the reset button on the front of the PC. The Pentium 4 can be configured to use the interrupts the same way as on the 8088 (for purposes of backward compatibility) or it can also use a new interrupt system using a device called an APIC (Advanced Programmable Interrupt Controller). The Pentium 4 can run at any one of several predefined voltages, but it has to know which one. The power management signals are used for automatic power supply voltage selection, telling the CPU that power is stable, and other power-

194

THE DIGITAL LOGIC LEVEL

CHAP. 3

related matters. Managing the various sleep states is also done here since sleeping is done for reasons of power management. Despite sophisticated power management, the Pentium 4 can get very hot. The Heat management group deals with thermal management, allowing the CPU to indicate to its environment that it is in danger of overheating. One of the pins is asserted by the CPU if its internal temperature reaches 130°C (266°F). If a CPU ever hits this temperature, it is probably dreaming about retirement and becoming a camp stove. The Clock frequency group has to do with determining the frequency of the system bus. The Diagnostics group contains signals for testing and debugging systems in conformance with the IEEE 1149.1 JTAG test standard. The Initialization group deals with booting (starting) the system. Finally, the miscellaneous group is a hodge-podge of signals including one that indicates if the CPU socket is occupied, one that relates to 8088 emulation, and other signals that have various special purposes. Pipelining on the Pentium 4’s Memory Bus Modern CPUs like the Pentium 4 are much faster than modern DRAM memories. To keep the CPU from starving for lack of data, it is essential to get the maximum possible throughput from the memory. For this reason, the Pentium 4 memory bus is highly pipelined, with as many as eight bus transactions going on at the same time. We saw the concept of pipelining in Chap. 2 in the context of a pipelined CPU (see Fig. 2-4), but memories can also be pipelined. To allow pipelining, Pentium 4 memory requests, called transactions, have six stages: 1. The bus arbitration phase. 2. The request phase. 3. The error-reporting phase. 4. The snoop phase. 5. The response phase. 6. The data phase. Not all phases are needed on all transactions. The bus arbitration phase determines which of the potential bus masters goes next. The request phase allows the address to be put onto the bus and the request made. The error-reporting phase allows the slave to announce that the address had a parity error or that something else is wrong. The snoop phase allows one CPU to snoop on the other one, something only needed in a multiprocessor system. The response phase is where the master learns about whether it is about to get the data it wants. Finally, the data phase allows the data to be sent back to the CPU requesting it.

SEC. 3.5

195

EXAMPLE CPU CHIPS

The secret to the Pentium 4’s pipelined memory bus is that each phase uses different bus signals, so that each one is completely independent of the other ones. The six groups of signals needed are the ones shown in Fig. 3-45 on the left. For example, one CPU can try to get the bus using the arbitration signals. Once it has acquired the right to go next, it releases these bus lines and starts using the Request group’s lines. Meanwhile, the other CPU or some I/O device can enter the bus arbitration phase, and so on. Figure 3-46 shows how multiple bus transactions can be outstanding at the same time. Bus cycle T1

T2

T3

T4

T5

T6

T7

T8

T9

Req

Error

Snoop

Resp

Data

Req

Error

Snoop

Resp

Data

Req

Error

Snoop

Resp

Req

Error

Snoop

Resp

Req

Error

Snoop

Req

Error

Snoop

Req

Error

T10

T11

T12

Φ Transaction 1 2 3 4 5 6 7

Data Data Resp

Data Resp

Snoop

Data Resp

Data

Figure 3-46. Pipelining requests on the Pentium 4’s memory bus.

In Fig. 3-46 the bus arbitration phase is not shown because it is not always needed. For example, if the current bus owner (often the CPU) wants to run another transaction, it does not have to reacquire the bus. It only has to ask for the bus again after it passes bus ownership to another requesting device. Transactions 1 and 2 are straightforward: five phases in five bus cycles. Transaction 3 introduces a longer data phase, for example, because it is a block transfer or because the memory addressed inserted a wait state. As a consequence, transaction 4 cannot start its data phase when it would like to. It observes that the DBSY# signal is still asserted and just waits for it to be negated. In transaction 5 we see that the response phase can also take multiple bus cycles, thus delaying transaction 6. Finally, in transaction 7, we notice that once a bubble has been introduced into the pipeline it remains there if new transactions keep starting consecutively. In actual practice, however, it is unlikely that the CPU will attempt to start a new transaction on every single bus cycle, so bubbles do not last so long.

196

THE DIGITAL LOGIC LEVEL

CHAP. 3

3.5.2 The UltraSPARC III As our second example of a CPU chip, we will now examine the Sun UltraSPARC family. The UltraSPARC family is Sun’s line of 64-bit SPARC CPUs. These CPUs fully conform to the Version 9 SPARC architecture, which is also for 64-bit CPUs. They are used in Sun workstations and servers, as well as various other applications. This family includes the UltraSPARC I, UltraSPARC II, and UltraSPARC III, which are architecturally very similar, differing primarily in introduction date, clock speed, and a few extra instructions introduced with each model. For the sake of concreteness, we will refer to the UltraSPARC III below, but most of the architectural (i.e., technology-independent) discussion holds for the other UltraSPARCs as well. The UltraSPARC III is a traditional RISC machine and is fully binary compatible with the 32-bit SPARC V8 architecture. It can run 32-bit SPARC V8 binary programs without modification because the SPARC V9 architecture is backward compatible with the SPARC V8 architecture. The only place where the UltraSPARC III deviates from the SPARC V9 architecture is the addition of the VIS 2.0 instruction set, which is designed for 3D graphical applications, real-time MPEG decoding, data compression, signal processing, running Java programs, and networking. Although the UltraSPARC III is also used in workstations, it was really designed to be used in Sun’s core business, large shared-memory multiprocessor servers used on the Internet and on corporate intranets. Specifically, much of the ‘‘glue’’ needed to build a multiprocessor is included inside each UltraSPARC III chip, making it easier to connect large numbers of them together. The first UltraSPARC III was introduced in 2000 at 600 MHz using 0.18 micron aluminum lines. The chips contain 29 million transistors. Because Sun’s volume is too small to warrant building a state-of-the-art chip fabrication plant, it prefers to concentrate on chip design and software and contract CPU manufacturing out to chip vendors. In the case of the UltraSPARC III, the chips are manufactured by Texas Instruments. In 2001, TI improved its technology and began making 900 MHz 0.15 micron chips using copper instead of aluminum wires. In 2002, the line width dropped to 0.13 microns and the clock was raised to 1.2 GHz. These chips require 50 watts of power and thus have about the same heat dissipation problems as the Pentium 4. It is difficult to compare a CISC chip (like the Pentium 4) and a RISC chip (like the UltraSPARC III) based on clock speed alone. For example, the UltraSPARC III can continuously issue four instructions per clock cycle, giving it almost the same execution rate as a single-issue CPU running at 4.8 GHz. The UltraSPARC also has six internal pipelines, including two 14-stage pipelines for integer operations, two for floating-point operations, one for load/store operations, and one for branches. It also has a different approach to caching, wider buses, and other factors that improve performance. The Pentium 4 also has its particular

SEC. 3.5

EXAMPLE CPU CHIPS

197

strengths. The point here is that just comparing two very different chips based on their relative clock speeds says very little about relative performance at some specific task. The UltraSPARC III comes in a 1368-pin Land Grid Array, as shown in Fig. 3-47. This package consists of a square array of 37 × 37 = 1369 pins on the bottom of the chip, with the pin in the lower left-hand corner missing. The socket exactly matches the chip to prevent the chip from being inserted incorrectly in the socket.

Figure 3-47. The UltraSPARC III CPU chip.

The UltraSPARC III has two main internal L1 caches: 32 KB for instructions and 64 KB for data. There is also a 2-KB prefetch cache and a 2-KB write cache used to collect writes to the level 2 cache so they can be done in large bursts to improve bandwidth usage. Like the Pentium 4, it also uses an off-chip level 2 cache, but unlike the Pentium 4, the UltraSPARC III is not packaged with the level 2 cache on chip. The cache controller and logic for locating cache blocks is on chip, but the actual SRAM memory is not. Instead, system designers are free to choose any commercially-available cache chips they want for the level 2 cache. The decision to integrate the level 2 cache on the Pentium 4 and separate it on the UltraSPARC III is partly due to technical issues and partly due to different business models used by Intel and Sun. On the technical side, an external cache is larger and more flexible (UltraSPARC III L2 caches can range from 1 MB to 8 MB; Pentium 4 L2 caches are fixed at 512 KB). However, it may be slower due to its greater distance from the CPU. It also requires more visible signals to address the cache. In particular, the connection between the UltraSPARC III and its L2 caches is 256 bits wide, allowing for an entire 32-byte cache block to be transferred in one cycles. On the business side, Intel is a semiconductor vendor and has the capability to design and manufacture its own level 2 cache chip and connect it to the CPU via a high-performance proprietary interface. Sun, in contrast, makes computers, not chips. It does design some of its own chips (like the UltraSPARCs), but farms out

198

THE DIGITAL LOGIC LEVEL

CHAP. 3

the manufacturing to semiconductor manufacturers. When it can, Sun prefers to use commercially available chips that have been keenly honed by the competitive marketplace. The SRAMs used for level 2 caches are available from numerous chip vendors, so there was no special need for Sun to design its own. This decision implies making the level 2 cache independent of the CPU chip. The UltraSPARC III uses a 43-bit wide address bus, allowing it to have up to 8 TB of main memory. The data bus is 128 bits wide, allowing 16 bytes at a time to be transferred between the CPU and memory. The bus speed is 150 MHz, giving a memory bandwidth of 2.4 GB/sec, much faster than the 528 MB/sec of the PCI bus. To connect (multiple) UltraSPARC CPUs to communicate with (multiple) memories, Sun developed the UPA (Ultra Port Architecture). The UPA can be implemented as a bus, a switch, or a combination of the two. Different workstation and server models use different UPA implementations. The UPA implementation does not matter to the CPU because the interface to the UPA is precisely defined, and it is this interface that the CPU chip must (and does) support. In Fig. 3-48 we see the core of an UltraSPARC III system, showing the CPU chip, UPA interface, and level 2 cache (two commodity SRAMs). The figure also contains a UDB II (UltraSPARC Data Buffer II) chip, whose function will be discussed below. When the CPU needs a memory word, it first looks in one of its (internal) level 1 caches for it. If it finds the word, it continues execution at full speed. If it does not find the word in the level 1 cache, it tries the level 2 cache. While we will discuss caching in detail in Chap. 4, a few words about it here will be useful. All of main memory is divided up into cache lines (blocks) of 64 bytes. The 256 most heavily used instruction lines and the 256 most heavily used data lines are in the level 1 cache. Cache lines that are heavily used but which do not fit in the level 1 cache are kept in the level 2 cache. This cache contains both data lines and instruction lines mixed at random. These are stored in the rectangle labeled ‘‘Level 2 cache data.’’ The system has to keep track of which lines are in the level 2 cache. This information is kept in a second SRAM, labeled ‘‘Level 2 cache tags.’’ On a level 1 cache miss, the CPU sends the identifier of the line it is looking for (Tag address) to the level 2 cache. The reply (Tag data) provides the information for the CPU to tell whether the line is in the level 2 cache, and if so, what state it is in. If the line is cached there, the CPU goes and gets it. The data transfers are 16 bytes wide, so four cycles are needed to fetch an entire line into the level 1 cache. If the cache line is not in the level 2 cache, it must be fetched from main memory via the UPA interface. The UltraSPARC III UPA is implemented with a centralized controller. The address and control signals from the CPU (all the CPUs if there are more than one of them) go there. To access memory, the CPU must first use the bus arbitration pins to acquire permission to go next. Once permission has been granted, the CPU outputs the memory address pins, specifies the

SEC. 3.5 18

Tag address Tag valid

Level 2 cache tags

Bus arbitration

5

Memory address

35

Address parity 25

Tag data

4

Tag parity

20

Address valid UltraSPARC II CPU

Data address

Wait Reply

Data address valid Level 2 cache data

199

EXAMPLE CPU CHIPS

UPA interface to main memory

4

Level 1 caches 128

Data

16

Parity

5 Control UDB II memory buffer

Memory data

128

Memory ECC

16

Figure 3-48. The main features of the core of an UltraSPARC III system.

request type, and asserts the address valid pin. (These pins are bidirectional since other CPUs in a UltraSPARC III multiprocessor need access to remote caches to keep all the caches consistent.) The address and bus cycle type are put out in two cycles on the Address pins, with the row going out in the first cycle and the column going out in the second cycle, as we saw in Fig. 3-31. While waiting for the results, the CPU may well be able to continue with other work. For example, a cache miss while prefetching an instruction does not inhibit the execution of one or more instructions already fetched, each of which may refer to data not in any cache. Thus multiple transactions to the UPA may be outstanding at once. The UPA can handle two independent transaction streams (typically reads and writes), each with multiple transactions pending. It is up to the centralized controller to keep track of all this and to make actual memory requests in the most efficient order. When data finally arrives from the memory, it can come in 8 bytes at a time, and with a 16-bit error-correcting code for greater reliability. A transaction may ask for an entire cache block, a quadword (8 bytes), or even fewer bytes. All incoming data go to the UDB, which buffers them. The purpose of the UDB is to further decouple the CPU from the memory system, so they can work asynchronously. For example, if the CPU has to write a word or cache line to memory, instead of waiting to access the UPA, it can write the data to the UDB immediately and let the UDB handle getting them to memory later. The UDB also generates

200

THE DIGITAL LOGIC LEVEL

CHAP. 3

and checks the error-correcting code. Just for the record, the description of the UltraSPARC III given above, like the one of the Pentium 4 before it, has been highly simplified, but the essence of its operation has been described.

3.5.3 The 8051 Both the Pentium 4 and the UltraSPARC III are examples of highperformance CPUs designed for building extremely fast PCs and servers. When many people think about computers, this is the kind of system they tend to focus on. However, there is another whole world of computers that is actually far larger: embedded systems. In this section we will take a brief look at that world. It is probably only a slight exaggeration to say that every electrical device costing more than 100 dollars has a computer in it. Certainly televisions, cell phones, electronic personal organizers, microwave ovens, camcorders, VCRs, laser printers, burglar alarms, hearing aids, electronic games, and other devices too numerous to mention are all computer controlled these days. The computers inside these things tend to be optimized for low price rather than for high performance, which leads to different trade-offs than the high-end CPUs we have been studying so far. As we mentioned in Chap. 1, the 8051 is probably the most popular microcontroller in current use, mostly due to its very low cost. As we will see shortly, it is also a simple chip, which makes interfacing to it simple and inexpensive. So let us now examine the 8051 chip, whose physical pinout is shown in Fig. 3-49. P1.0 P1.1 P1.2 P1.3 P1.4 P1.5 P1.6 P1.7 RST RXD / P3.0 TXD / P3.1 INT0 / P3.2 INT1 / P3.3 TO / P3.4 T1 / P3.5 WR / P3.6 RD / P3.7 XTAL2 XTAL1 VSS

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

40 39 38 37 36 35 34 33 32 31 30 29 28 27 26 25 24 23 22 21

VCC P0.0 / AD0 P0.1 / AD1 P0.2 / AD2 P0.3 / AD3 P0.4 / AD4 P0.5 / AD5 P0.6 / AD6 P0.7 / AD7 EA / VPP ALE PSEN P2.7 / A15 P2.6 / A14 P2.5 / A13 P2.4 / A12 P2.3 / A11 P2.2 / A10 P2.1 / A9 P2.0 / A8

Figure 3-49. Physical pinout of the 8051.

As can be seen from the figure, the 8051 normally comes in a standard 40-pin package (although other packages are available for special uses). It has 16

SEC. 3.5

201

EXAMPLE CPU CHIPS

address lines, so it can address up to 64 KB of memory. The data bus is 8 bits wide, so data transfers between the CPU and memory are done one byte at a time (versus 8 bytes at a time on the Pentium 4 and 16 bytes at a time on the UltraSPARC III). It has a variety of control lines, described below, but the greatest contrast with the Pentium 4 and UltraSPARC, which are pure CPUs, is the presence of 32 I/O lines, arranged in four groups of 8 bits each. Each of these I/O lines can be attached to a button, switch, LED (Light Emitting Diode), or other real-world device to provide input to the 8051 or output from the 8051. For example, in a clock radio, each of the buttons and switches could be wired to a different I/O line, with other I/O lines controlling the display. In this way, most, if not all, the functions of the clock radio could be controlled in software, eliminating the need for costly discrete logic. The logical pinout of the 8051 is shown in Fig. 3-50. The 8051 comes with 4 KB of internal ROM (8 KB on the 8052). If that is insufficient for the application, up to 64 KB of external memory may be connected to the 8051 over a bus. The first seven signals on the left-hand side of Fig. 3-50 are used to interface to external memories, if present. The first signal, A, contains 16 address lines to address the byte of external memory to be read or written. The eight D lines are used for data transport. The low-order eight address lines are multiplexed onto the same pins as the data lines to reduce pin count. On a bus transaction, these pins output the address on the first clock cycle and carry the data on subsequent cycles. A D RD WR ALE PSEN EA Timers Interrupts

16 8

8

8

2

Port 0

Port 1

8051 8

2

TXD RXD

8

Port 2

Port 3

RST 2 Φ Power

Figure 3-50. Logical pinout of the 8051.

indicate whether it is readWhen an external memory is used, the 8051 """has to""" ing or writing memory by asserting either RD or WR, respectively. The ALE

202

THE DIGITAL LOGIC LEVEL

CHAP. 3

(Address Latch Enable) signal is used when an external memory is present. The CPU asserts this signal to indicate that the address is valid. External memories typically use it to latch the address lines, since they will be released shortly thereafter""""" so the pins """"can be reused for the data. """"" The PSEN and EA# signals also relate to external memory. The PSEN (Program Store ENable) signal is asserted to indicate that the 8051 wants to read from the program memory. Typically, it would be connected to the memory’s OE signal, as illustrated in Fig. 3-29. """" The EA# (External Access) signal is usually wired high or low so it always has the same value. If it is wired high, the internal 4 KB (8 KB on the 8052) memory is used for addresses within range and external memory is used for addresses above 4 KB (8 KB on the 8052). If it is wired low, the external memory is used for all addresses and the on-chip memory is effectively bypassed. On the 8031 """" and 8032, EA# must be wired low because there is no on-chip memory. The two timer lines allow external timers to be input to the CPU. The two interrupt lines allow two different external devices to interrupt the CPU. The TXD and RXD lines are to allow serial I/O to a terminal or modem. Finally, the RST line allows the user or external hardware device to reset the 8051. This line is typically asserted when something has gone wrong and the system has to be rebooted. So, far the 8051 is similar to most other 8-bit CPUs, except for the presence of the serial I/O lines. What sets the 8051 apart is the presence of 32 I/O lines, organized as four ports and shown on the right-hand side of Fig. 3-50. Each of these lines is bidirectional and can be read or written under program control. This is the primary way the 8051 interacts with the outside world and what makes it so valuable: a single chip is all that is needed for CPU, memory, and I/O capability.

3.6 EXAMPLE BUSES Buses are the glue that hold computer systems together. In this section we will take a close look at some popular buses: the ISA bus, the PCI bus, and the Universal Serial Bus. The ISA bus is a slight expansion of the original IBM PC bus. For reasons of backward compatibility, it was still present in all Intel-based PCS until a few years ago when Intel and Microsoft agreed to eliminate it. However, these machines invariably have a second, faster bus as well: the PCI bus. The PCI bus is wider than the ISA bus and runs at a higher clock rate. Consequently it can carry more data per second than the ISA bus. It is the workhorse for most current PCs, although a successor is already in sight. The Universal Serial Bus is an increasingly popular I/O bus for low-speed peripherals such as mice and keyboards. A second version of the USB bus runs at much higher speeds. In the following sections, we will look at each of these buses in turn to see how they work.

SEC. 3.6

EXAMPLE BUSES

203

3.6.1 The ISA Bus The IBM PC bus was the de facto standard on 8088-based systems because nearly all PC clone vendors copied it in order to allow the many existing thirdparty I/O boards to be used with their systems. It had 62 signal lines, including 20 for a memory address, 8 for data, and one each for asserting memory read, memory write, I/O read, and I/O write. There were also signals for requesting and granting interrupts and using DMA, and that was about it. It was a very simple bus. Physically, the bus was etched onto the PC’s motherboard, with about half a dozen connectors spaced 2 cm apart into which cards could be inserted. Each card had a tab on it that fit in the connector. The tab had 31 gold-plated strips on each side that made electrical contact with the connector. When IBM introduced the 80286-based PC/AT, it had a major problem on its hands. If it had started from scratch and designed an entirely new 16-bit bus, many potential customers would have hesitated to buy it because none of the vast number of PC plug-in boards available from third-party vendors would have worked using the new machine. On the other hand, sticking with the PC bus and its 20 address lines and 8 data lines would not have taken advantage of the 80286’s ability to address 16 MB of memory and transfer 16-bit words. The solution chosen was to extend the PC bus. PC plug-in cards have an edge connector with 62 contacts, but this edge connector does not run the full length of the board. The PC/AT solution was to put a second edge connector on the bottom of the board, adjacent to the main one, and design the AT circuitry to work with both types of boards. The general idea is illustrated in Fig. 3-51. The second connector on the PC/AT bus contains 36 lines. Of these, 31 are provided for more address lines, more data lines, more interrupt lines, and more DMA channels, as well as power and ground. The rest deal with differences between 8-bit and 16-bit transfers. When IBM brought out the PS/2 series as the successor to the PC and PC/AT, it decided that it was time to start again. Part of this decision may have been technical (the PC bus was by this time really obsolete), but part was, no doubt, caused by a desire to put an obstacle in the way of companies making PC clones, which had taken over an uncomfortably large part of the market. Thus the midand upper-range PS/2 machines were equipped with a bus, the Microchannel, that was completely new, and which was protected by a wall of patents backed by an army of lawyers. The rest of the personal computer industry reacted to this development by adopting its own standard, the ISA (Industry Standard Architecture) bus, which was basically the PC/AT bus running at 8.33 MHz. The big advantage of this approach was that it retained compatibility with existing machines and cards. It also was based on a bus that IBM had liberally licensed to many companies in order to ensure that as many third parties as possible produced cards for the

204

THE DIGITAL LOGIC LEVEL

Motherboard

PC bus connector

PC bus

CHAP. 3

Plug-in Contact board Chips

CPU and other chips

New connector for PC/AT

Edge connector

Figure 3-51. The PC/AT bus has two components, the original PC part and the new part.

original PC, something that came back to haunt IBM and ultimately drive it from the PC business. Until a few years ago, most Intel-based PCs still had this bus present, although with one or more other buses as well. Later, the ISA bus was extended to 32 bits with some new features thrown in (e.g., for multiprocessing).The new bus was called the EISA (Extended ISA) bus.

3.6.2 The PCI Bus On the original IBM PC, most applications were text based. Gradually, with the introduction of Windows, graphical user interfaces came into use. None of these applications put much of a strain on the ISA bus. However, as time went on and many applications, especially multimedia games, began to use computers to display full-screen, full-motion video, the situation changed radically. Let us make a simple calculation. Consider a 1024 × 768 screen used for true color (3 bytes/pixel) moving images. One screen contains 2.25 MB of data. For smooth motion, at least 30 screens/sec are needed, for a data rate of 67.5 MB/sec. In fact, it is worse than this, since to display a video from a hard disk, CD-ROM, or DVD, the data must pass from the disk drive over the bus to the memory. Then for the display, the data must travel over the bus again to the graphics adapter. Thus we need a bus bandwidth of 135 MB/sec for the video alone, not counting the bandwidth the CPU and other devices need.

SEC. 3.6

EXAMPLE BUSES

205

The ISA bus ran at a maximum rate of 8.33 MHz, and could transfer 2 bytes per cycle, for a maximum bandwidth of 16.7 MB/sec. The EISA bus could move 4 bytes per cycle, to achieve 33.3 MB/sec. Clearly, neither of these was even close to what is needed for full-screen video. In 1990, Intel saw this coming and designed a new bus with a much higher bandwidth than even the EISA bus. It was called the PCI bus (Peripheral Component Interconnect bus). To encourage its use, Intel patented the PCI bus and then put all the patents into the public domain, so any company could build peripherals for it without having to pay royalties. Intel also formed an industry consortium, the PCI Special Interest Group, to manage the future of the PCI bus. As a result of these actions, the PCI bus became extremely popular. Virtually every Intel-based computer since the Pentium has a PCI bus, and many other computers do, too. Sun even has a version of the UltraSPARC that uses the PCI bus, the UltraSPARC IIIi. The PCI bus is covered in gory detail in Shanley and Anderson (1999) and Solari and Willse (2004). The original PCI bus transferred 32 bits per cycle and ran at 33 MHz (30 nsec cycle time) for a total bandwidth of 133 MB/sec. In 1993, PCI 2.0 was introduced, and in 1995, PCI 2.1 came out. PCI 2.2 has features for mobile computers (mostly for saving battery power). The PCI bus runs at up to 66 MHz and can handle 64-bit transfers, for a total bandwidth of 528 MB/sec. With this kind of capacity, full-screen, full-motion video is doable (assuming the disk and the rest of the system are up to the job). In any event, the PCI bus will not be the bottleneck. Even though 528 MB/sec sounds pretty fast, it still had two problems. First, it was not good enough for a memory bus. Second, it was not compatible with all those old ISA cards out there. The solution Intel thought of was to design computers with three or more buses, as shown in Fig. 3-52. Here we see that the CPU can talk to the main memory on a special memory bus, and that an ISA bus can be connected to the PCI bus. This arrangement met all requirements, and as a consequence was widely used in the 1990s. Two key components in this architecture are the two bridge chips (which Intel manufactures—hence its interest in this whole project). The PCI bridge connects the CPU, memory and PCI bus. The ISA bridge connects the PCI bus to the ISA bus and also supports one or two IDE disks. Nearly all Pentium 4 systems come with one or more free PCI slots for adding new high-speed peripherals, and one or more ISA slots, for adding low-speed peripherals. The big advantage of the design of Fig. 3-52 is that the CPU has an extremely high bandwidth to memory using a proprietary memory bus; the PCI bus offers high bandwidth for fast peripherals such as SCSI disks, graphics adaptors, etc.; and old ISA cards can still be used. The USB box in the figure refers to the Universal Serial Bus, which will be discussed later in this chapter. It would have been nice had there been only one kind of PCI card. Unfortunately, such is not the case. Options are provided for voltage, width, and timing.

206

THE DIGITAL LOGIC LEVEL Cache bus

Level 2 cache

CHAP. 3

Memory bus

Local bus PCI bridge

CPU

Main memory PCI bus

SCSI

USB

Mouse

Modem

ISA bridge

IDE disk

Graphics adaptor

Available PCI slot

Monitor

Keyboard

ISA bus

Sound card

Printer

Available ISA slot

Figure 3-52. Architecture of an early Pentium system. The thicker buses have more bandwidth than the thinner ones but the figure is not to scale.

Older computers often use 5 volts and newer ones tend to use 3.3 volts, so the PCI bus supports both. The connectors are the same except for two bits of plastic that are there to prevent people from inserting a 5-volt card in a 3.3-volt PCI bus or vice versa. Fortunately, universal cards, which support both voltages and can plug into either kind of slot, exist. In addition to the voltage option, cards come in 32-bit and 64-bit versions. The 32-bit cards have 120 pins; the 64-bit cards have the same 120 pins plus an additional 64 pins, analogous to the way the IBM PC bus was extended to 16 bits (see Fig. 3-51). A PCI bus system that supports 64bit cards can also take 32-bit cards, but the reverse is not true. Finally, PCI buses and cards can run at either 33 MHz or 66 MHz. The choice is made by having one pin wired either to the power supply or wired to ground. The connectors are identical for both speeds. By the late 1990s, pretty much everyone agreed that the ISA bus was dead, so new designs excluded it. However, by then, monitor resolution had increased in some cases to 1600 × 1200 and the demand for full-screen full motion video had also increased, especially in the context of highly interactive games, so Intel added yet another bus just to drive the graphics card. This bus was called the AGP bus (Accelerated Graphics Port bus). The initial version, AGP 1.0, ran at

SEC. 3.6

207

EXAMPLE BUSES

264 MB/sec, which was defined as 1x. While slower than the PCI bus, it was dedicated to driving the graphics card. Over the years, new versions came out, with AGP 3.0 running at 2.1 GB/sec (8x). A modern Pentium 4 system is illustrated in Fig. 3-53. Monitor

Graphics adaptor AGP bus Memory bus

Local bus

Level 1 caches Level 2 cache

Pentium 4 CPU I

Bridge chip

D

Main memory

PCI bus

SCSI

USB 2

Mouse

Keyboard

Available PCI slot

ATAPI controller

Hard disk

DVD drive

Figure 3-53. The bus structure of a modern Pentium 4.

In this design, the bridge chip is now central. It connects the five major pieces of the system: the CPU, memory, the graphics card, the ATAPI controller, and the PCI bus. In some variations, it also has support for Ethernet and other high-speed devices. The lower-speed devices are attached to the PCI bus. Internally, the bridge chip is divided into two parts: the memory bridge and the I/O bridge. The memory bridge connects the CPU to the memory and to the graphics adaptor. The I/O bridge connects the ATAPI controller, PCI bus, and (optionally) other fast I/O devices with a direct bridge connection to each other. The two bridges are connected by a very high speed interconnect. The PCI bus is synchronous, like all PC buses going back to the original IBM PC. All transactions on the PCI bus are between a master, officially called the initiator, and a slave, officially called the target. To keep the PCI pin count down, the address and data lines are multiplexed. In this way, only 64 pins are

208

THE DIGITAL LOGIC LEVEL

CHAP. 3

needed on PCI cards for address plus data signals, even though PCI supports 64bit addresses and 64-bit data. The multiplexed address and data pins work as follows. On a read operation, during cycle 1, the master puts the address onto the bus. On cycle 2, the master removes the address and the bus is turned around so the slave can use it. On cycle 3, the slave outputs the data requested. On write operations, the bus does not have to be turned around because the master puts on both the address and the data. Nevertheless, the minimum transaction is still three cycles. If the slave is not able to respond in three cycles, it can insert wait states. Block transfers of unlimited size are also allowed, as well as several other kinds of bus cycles. PCI Bus Arbitration

PCI device

PCI device

PCI device

GNT#

REQ#

GNT#

REQ#

GNT#

REQ#

GNT#

PCI arbiter

REQ#

To use the PCI bus, a device must first acquire it. PCI bus arbitration uses a centralized bus arbiter, as shown in Fig. 3-54. In most designs, the bus arbiter is built into one of the bridge chips. Every PCI device has two dedicated lines running from it to the arbiter. One line, REQ#, is used to request the bus. The other line, GNT#, is used to receive bus grants.

PCI device

Figure 3-54. The PCI bus uses a centralized bus arbiter.

To request the bus, a PCI device (including the CPU), asserts REQ# and waits until it sees its GNT# line asserted by the arbiter. When that event happens, the device can use the bus on the next cycle. The algorithm used by the arbiter is not defined by the PCI specification. Round-robin arbitration, priority arbitration, and other schemes are all allowed. Clearly, a good arbiter will be fair, so as not to let some devices wait forever. A bus grant is for only one transaction, although the length of this transaction is theoretically unbounded. If a device wants to run a second transaction and no other device is requesting the bus, it can go again, although often one idle cycle between transactions has to be inserted. However, under special circumstances, in

SEC. 3.6

EXAMPLE BUSES

209

the absence of competition for the bus, a device can make back-to-back transactions without having to insert an idle cycle. If a bus master is making a very long transfer and some other device has requested the bus, the arbiter can negate the GNT# line. The current bus master is expected to monitor the GNT# line, so when it sees the negation, it must release the bus on the next cycle. This scheme allows very long transfers (which are efficient) when there is only one candidate bus master but still gives fast response to competing devices. PCI Bus Signals The PCI bus has a number of mandatory signals, shown in Fig. 3-55(a), and a number of optional signals, shown in Fig. 3-55(b). The remainder of the 120 or 184 pins are used for power, ground, and related miscellaneous functions and are not listed here. The Master (initiator) and Slave (target) columns tell who asserts the signal on a normal transaction. If the signal is asserted by a different device (e.g., CLK), both columns are left blank. Let us now look at each of the PCI bus signals briefly. We will start with the mandatory (32-bit) signals; then we will move on to the optional (64-bit) signals. The CLK signal drives the bus. Most of the other signals are synchronous with it. In contrast to the ISA bus, a PCI bus transaction begins at the falling edge of CLK, which is in the middle of the cycle, rather than at the start. The 32 AD signals are for the address and data (for 32-bit transactions). Generally, during cycle 1 the address is asserted and during cycle 3 the data are asserted. The PAR signal is a parity bit for AD. The C/BE# signal is used for two different things. On cycle 1, it contains the bus command (read 1 word, block read, etc.). On cycle 2 it contains a bit map of 4 bits, telling which bytes of the 32-bit word are valid. Using C/BE# it is possible to read or write any 1, 2, or 3 bytes, as well as an entire word. The FRAME# signal is asserted by the bus master to start a bus transaction. It tells the slave that the address and bus commands are now valid. On a read, usually IRDY# is asserted at the same time as FRAME#. It says the master is ready to accept incoming data. On a write, IRDY# is asserted later, when the data are on the bus. The IDSEL signal relates to the fact that every PCI device must have a 256byte configuration space that other devices can read (by asserting IDSEL ). This configuration space contains properties of the device. The Plug-and-Play feature of some operating systems uses the configuration space to find out what devices are on the bus. Now we come to signals asserted by the slave. The first of these, DEVSEL# , announces that the slave has detected its address on the AD lines and is prepared to engage in the transaction. If DEVSEL# is not asserted within a certain time limit, the master times out and assumes the device addressed is either absent or broken.

210

THE DIGITAL LOGIC LEVEL

CHAP. 3

################################################################################ ! Signal ! Lines ! Master ! Slave ! Description ! ! ! ! ! !################################################################################ CLK 1 ! ! ! ! Clock (33 MHz or 66 MHz) ! !################################################################################ ! 32 ! ! ! Multiplexed address and data lines ! ! AD × × ! ! ! ! ! !################################################################################ ! ! ! Address or data parity bit ! !################################################################################ PAR 1 ! × ! ! ! ! ! ! C/BE 4 × Bus command/bit map for bytes enabled ! ! ! ! ! !################################################################################ ! ! Indicates that AD and C/BE are asserted ! !################################################################################ FRAME# ! 1 ! × ! ! ! ! ! ! IRDY# 1 ! × ! ! ! Read: master will accept; write: data present ! !################################################################################ ! ! ! ! Select configuration space instead of memory ! ! IDSEL 1 × ! ! ! ! ! !################################################################################ DEVSEL# ! 1 ! × ! ! Slave has decoded its address and is listening ! !################################################################################ ! ! ! Read: data present; write: slave will accept ! ! TRDY# 1 ! × ! ! ! ! ! !################################################################################ ! ! ! Slave wants to stop transaction immediately ! !################################################################################ STOP# 1 ! × ! ! ! ! ! ! PERR# 1 ! ! ! Data parity error detected by receiver ! ! !################################################################################ ! ! Address parity error or system error detected ! ! ! SERR# 1 ! # ############################################################################### ! ! ! ! ! ! REQ# 1 ! ! ! Bus arbitration: request for bus ownership ! ! !################################################################################ ! ! Bus arbitration: grant of bus ownership ! ! ! GNT# 1 ! # ############################################################################### ! ! ! ! ! ! 1 !! RST# !! !! Reset the system and all devices !! !! !!################################################################################ (a) !

################################################################################ ! Signal ! Lines ! Master ! Slave ! ! Description ################################################################################ ! ! ! ! ! ! 1 ! × ! REQ64# ! ! ! Request to run a 64-bit transaction ! ################################################################################ ! ACK64# ! ! ! ! Permission is granted for a 64-bit transaction ! 1 × ################################################################################ ! ! ! ! ! ! ! AD ! 32 ! ! ! Additional 32 bits of address or data ! × ################################################################################ ! ! ! ! ! ! PAR64 1 × Parity for the extra 32 address/data bits ################################################################################ ! ! ! ! ! ! !################################################################################ ! ! ! Additional 4 bits for byte enables ! C/BE# 4 ! × ! ! ! ! ! ! 1 ! × ! LOCK ! ! ! Lock the bus to allow multiple transactions ! ################################################################################ ! SBO# ! ! ! ! Hit on a remote cache (for a multiprocessor) ! 1 ################################################################################ ! ! ! ! ! ! 1 ! ! SDONE ! ! ! Snooping done (for a multiprocessor) ! ################################################################################ ! INTx ! ! ! ! Request an interrupt ! 4 ################################################################################ ! ! ! ! ! ! ! JTAG ! ! ! IEEE 1149.1 JTAG test signals ! 5 ! ################################################################################ ! ! ! ! ! ! M66EN 1 Wired to power or ground (66 MHz or 33 MHz) ! ################################################################################ ! ! ! ! ! (b) Figure 3-55. (a) Mandatory PCI bus signals. (b) Optional PCI bus signals.

The second slave signal is TRDY#, which the slave asserts on reads to announce that the data are on the AD lines and on writes to announce that it is prepared to accept data. The next three signals are for error reporting. The first of these is STOP#, which the slave asserts if something disastrous happens and it wants to abort the

SEC. 3.6

EXAMPLE BUSES

211

current transaction. The next one, PERR#, is used to report a data parity error on the previous cycle. For a read, it is asserted by the master; for a write it is asserted by the slave. It is up to the receiver to take the appropriate action. Finally, SERR# is for reporting address errors and system errors. The REQ# and GNT# signals are for doing bus arbitration. These are not asserted by the current bus master, but rather by a device that wants to become bus master. The last mandatory signal is RST#, used for resetting the system, either due to the user pushing the RESET button or some system device noticing a fatal error. Asserting this signal resets all devices and reboots the computer. Now we come to the optional signals, most of which relate to the expansion from 32 bits to 64 bits. The REQ64# and ACK64# signals allow the master to ask permission to conduct a 64-bit transaction and allow the slave to accept, respectively. The AD, PAR64, and C/BE# signals are just extensions of the corresponding 32-bit signals. The next three signals are not related to 32 bits versus 64 bits, but to multiprocessor systems, something that PCI boards are not required to support. The LOCK signal allows the bus to be locked for multiple transactions. The next two relate to bus snooping to maintain cache coherence. The INTx signals are for requesting interrupts. A PCI card can have up to four separate logical devices on it, and each one can have its own interrupt request line. The JTAG signals are for the IEEE 1149.1 JTAG testing procedure. Finally, the M66EN signal is either wired high or wired low, to set the clock speed. It must not change during system operation. PCI Bus Transactions The PCI bus is really very simple (as buses go). To get a better feel for it, consider the timing diagram of Fig. 3-56. Here we see a read transaction, followed by an idle cycle, followed by a write transaction by the same bus master. When the falling edge of the clock happens during T1 , the master puts the memory address on AD and the bus command on C/BE#. It then asserts FRAME# to start the bus transaction. During T2 , the master floats the address bus to let it turn around in preparation for the slave to drive it during T3 . The master also changes C/BE# to indicate which bytes in the word addressed it wants to enable (i.e., read in). In T3 , the slave asserts DEVSEL# so the master knows it got the address and is planning to respond. It also puts the data on the AD lines and asserts TRDY# to tell the master that it has done so. If the slave was not able to respond so quickly, it would still assert DEVSEL# to announce its presence but keep TRDY# negated until it could get the data out there. This procedure would introduce one or more wait states. In this example (and often in reality), the next cycle is idle. Starting in T5 we see the same master initiating a write. It starts out by putting the address and

212

THE DIGITAL LOGIC LEVEL

CHAP. 3

Bus cycle Idle

Read T1

T2

T3

T4

Write T5

T6

T7

Φ Turnaround AD C/BE#

Address Read cmd

Data Enable

Address

Data

Write cmd

Enable

FRAME# IRDY# DEVSEL# TRDY#

Figure 3-56. Examples of 32-bit PCI bus transactions. The first three cycles are used for a read operation, then an idle cycle, and then three cycles for a write operation.

command onto the bus, as usual. Only now, in the second cycle it asserts the data. Since the same device is driving the AD lines, there is no need for a turnaround cycle. In T7 , the memory accepts the data.

3.6.3 PCI Express Although the PCI bus works adequately for most current applications, the need for greater I/O bandwidth is making a mess of the once-clean internal PC architecture. In Fig. 3-53, it is clear that the PCI bus is no longer the central element that holds the parts of the PC together. The bridge chip has taken over part of that role. The essence of the problem is that there are increasingly many I/O devices that are too fast for the PCI bus. Cranking up the clock frequency on the bus is not a good solution because then problems with bus skew, crosstalk between the wires, and capacitance effects just get worse. Every time an I/O device gets too

SEC. 3.6

EXAMPLE BUSES

213

fast for the PCI bus (like the graphics card, hard disk, network, etc.), Intel adds a new special port to the bridge chip to allow that device to bypass the PCI bus. Clearly, this is not a long-term solution either. Another problem with the PCI bus is that the cards are quite large. They do not fit in laptop computers and palmtop computers and manufacturers would like to produce even smaller devices. Also, some manufacturers would like to repartition the PC, with the CPU and memory in a tiny sealed box and the hard disk inside the monitor. With PCI cards, doing this is impossible. Several solutions have been proposed, but the one most likely to win out (in no small part because Intel is behind it) is called PCI Express. It has little to do with the PCI bus, and in fact, is not a bus at all, but the marketing folks did not like letting go of the well-known PCI name. PCs containing it have been on the market for some time already. Let us now see how it works. The PCI Express Architecture The heart of the PCI Express solution is to get rid of the parallel bus with its many masters and slaves and go to a design based on high-speed point-to-point serial connections. This solution represents a radical break with the ISA/EISA/PCI bus tradition, and borrows many ideas from the world of local area networking, especially switched Ethernet. The basic idea comes down to this: deep inside, a PC is a collection of CPU, memory, and I/O controller chips that need to be interconnected. What PC Express does is provide a general-purpose switch for connecting chips using serial links. A typical configuration is illustrated in Fig. 3-57. As shown in Fig. 3-57, the CPU, memory, and cache are connected to the bridge chip in the traditional way. What is new here is a switch connected to the bridge (possibly part of the bridge chip itself). Each of the I/O chips has a dedicated point-to-point connection to the switch. Each connection consists of a pair of unidirectional channels, one to the switch and one from it. Each channel is made up of two wires, one for the signal and one for ground, to provide high noise immunity during high-speed transmission. This architecture will replace the current one with a much more uniform model, in which all devices are treated equally. The PCI Express architecture differs from the old PCI bus architecture in three key ways. We have already seen two of them: a centralized switch versus a multidrop bus and a the use of narrow serial point-to-point connections versus a wide parallel bus. The third one is more subtle. The conceptual model behind the PCI bus is that of a bus master issuing a command to a slave to read a word or a block of words. The PCI Express model is that of a device sending a data packet to another device. The concept of a packet, which consists of a header and a payload, is taken from the networking world. The header contains control information, thus eliminating the need for the many control signals present on the PCI

214

THE DIGITAL LOGIC LEVEL

Level 2 cache

CPU

Bridge chip

Memory

Switch

Graphics

Disk

Network

CHAP. 3

Paired serial links

USB 2

Other

Figure 3-57. A typical PCI Express system.

bus. The payload contains the data to be transferred. In effect, a PC with PCI Express is a miniature packet-switching network. In addition to these three major breaks with the past, there are also several minor differences as well. Fourth, an error-detecting code is used on the packets, providing a higher degree of reliability than on the PCI bus. Fifth, the connection between a chip and the switch is longer than it was, up to 50 cm, to allow system partitioning. Sixth, the system is expandable because a device may actually be another switch, allowing a tree of switches. Seventh, devices are hot pluggable, meaning that they can be added or removed from the system while it is running. Finally, since the serial connectors are much smaller than the old PCI connectors, devices and computers can be made much smaller. All in all, a major departure from the PCI bus. The PCI Express Protocol Stack In keeping with the model of a packet-switching network, the PCI Express system has a layered protocol stack. A protocol is a set of rules governing the conversation between two parties. A protocol stack is a hierarchy of protocols that deal with different issues at different layers. For example, consider a business letter. It has certain conventions about the placement and content of the letterhead, the recipient’s address, the date, the salutation, the body, the signature, and so on. This might be thought of as the letter protocol. In addition, there is

SEC. 3.6

215

EXAMPLE BUSES

another set of conventions about the envelope, such as its size, where the sender’s address goes and its format, where the receiver’s address goes and its format, where the stamp goes, and so on. These two layers and their protocols are independent. For example, it is possible to completely reformat the letter but use the same envelope or vice versa. Layered protocols make for a modular flexible design, and have been widely used in the world of network software for decades. What is new here is building them into the ‘‘bus’’ hardware. The PCI Express protocol stack is shown in Fig. 3-58(a). Software layer Header

Payload

Seq #

Header

Payload

CRC

Seq #

Header

Payload

CRC

Transaction layer Link layer Physical layer (a)

Frame

Frame

(b)

Figure 3-58. (a) The PCI Express protocol stack. (b) The format of a packet.

Let us examine the layers from the bottom up. The lowest layer is the physical layer. It deals with moving bits from a sender to a receiver over a point-topoint connection. Each point-to-point connection consists of one or more pairs of simplex (i.e., unidirectional) links. In the simplest case, there is one pair in each direction, but having 2, 4, 8, 16, or 32 pairs is also allowed. Each link is called a lane. The number of lanes in each direction must be the same. First-generation products must support a data rate each way of at least 2.5 Gbps, but the speed is expected to migrate to 10 Gbps each way fairly soon. Unlike the ISA/EISA/PCI buses, PCI Express does not have a master clock. Devices are free to start transmitting as soon as they have data to send. This freedom makes the system faster, but also leads to a problem. Suppose that a 1 bit is encoded as +3 volts and a 0 bit is encoded as 0 volts. If the first few bytes are all 0s, how does the receiver know data is being transmitted? After all a run of 0 bits looks the same as an idle link. The problem is solved using what is called 8b/10b encoding. In this scheme, 10 bits are used to encode 1 byte of actual data in a 10-bit symbol. Of the 1024 possible 10-bit symbols, the legal ones have been chosen to have enough clock transitions to keep the sender and receiver synchronized on the bit boundaries even without a master clock. A consequence of 8b/10b encoding is that a link with a gross capacity of 2.5 Gbps can carry only 2 Gbps of (net) user data. Whereas the physical layer deals with bit transmission, the link layer deals with packet transmission. It takes the header and payload given to it by the transaction layer and adds to them a sequence number and a error-correcting code called a CRC (Cyclic Redundancy Check). The CRC is generated by running a

216

THE DIGITAL LOGIC LEVEL

CHAP. 3

certain algorithm on the header and payload data. When a packet is received, the receiver performs the same computation on the header and data and compares the result with the CRC attached to the packet. If they agree, it sends back a short acknowledgment packet affirming its correct arrival. If they disagree, the receiver asks for a retransmission. In this way, data integrity is greatly improved over the PCI bus system, which does not have any provision for verification and retransmission of data sent over the bus. To prevent having a fast receiver bury a slow receiver in packets it cannot handle, a flow control mechanism is used. The mechanism used that the receiver gives the transmitter a certain number of credits, basically corresponding to the amount of buffer space it has available to store incoming packets. When the credits are used up, the transmitter has to stop sending until it is giving more credits. This scheme, which is widely used in all networks, prevents losing data due to a mismatch of transmitter and receiver speeds. The transaction layer handles bus actions. Reading a word from memory requires two transactions: one initiated by the CPU or DMA channel requesting some data and one initiated by the target supplying the data. But the transaction layer does more than handle pure reads and writes. It adds value to the raw packet transmission offered by the link layer. To start with, it can divide each lane into up to eight virtual circuits, each handling a different class of traffic. The transaction layer can tag packets according to their traffic class, which may include attributes such as high priority, low priority, do not snoop, may be delivered out of order, and more. The switch may use these tags when deciding which packet to handle next. Each transaction uses one of four address spaces: 1. Memory space (for ordinary reads and writes). 2. I/O space (for addressing device registers). 3. Configuration space (for system initialization, etc.). 4. Message space (for signaling, interrupts, etc.). The memory and I/O spaces are similar to what current systems have. The configuration space can be used to implement features such as plug-and-play. The message space takes over the role of many of the existing control signals. Something like this space is needed because none of the PCI bus’ control lines exist in PCI express. The software layer interfaces the PCI Express system to the operating system. It can emulate the PCI bus, making it possible to run existing operating systems unmodified on PCI Express systems. Of course, operating like this will not exploit the full power of PCI Express, but backward compatibility is a necessary evil that is needed until operating systems have been modified to fully utilize PCI Express. Experience shows that can take a while.

SEC. 3.6

EXAMPLE BUSES

217

The flow of information is illustrated in Fig. 3-58(b). When a command is given to the software layer, it hands it to the transaction layer, which formulates it in terms of a header and a payload. These two parts are then passed to the link layer, which attaches a sequence number to the front and a CRC to the back. This enlarged packet is then given to the physical layer, which adds framing information on each end to form the physical packet that is actually transmitted. At the receiving end, the reverse process takes place, with the link header and trailer being stripped and the result being given to the transaction layer. The concept of each layer adding additional information to the data as it works its way down the protocol has been used for decades in the networking world with great success. The big difference between a network and PCI Express is that in the networking world the code in the various layers is nearly always software that is part of the operating system. With PCI Express it is all part of the device hardware. PCI Express is a complicated subject. For more information see (Mayhew and Krishnan, 2003; and Solari and Congdon, 2005).

3.6.4 The Universal Serial Bus The PCI bus and PCI Express are fine for attaching high-speed peripherals to a computer, but they are too expensive for low-speed I/O devices such as keyboards and mice. Historically, each standard I/O device was connected to the computer in a special way, with some free ISA and PCI slots for adding new devices. Unfortunately, this scheme has been fraught with problems from the beginning. For example, each new I/O device often comes with its own ISA or PCI card. The user is often responsible for setting switches and jumpers on the card and making sure the settings do not conflict with other cards. Then the user must open up the case, carefully insert the card, close the case, and reboot the computer. For many users, this process is difficult and error prone. In addition, the number of ISA and PCI slots is very limited (two or three typically). Plug-and-play cards eliminate the jumper settings, but the user still has to open the computer to insert the card and bus slots are still limited. To deal with this problem, in 1993, representatives from seven companies (Compaq, DEC, IBM, Intel, Microsoft, NEC, and Northern Telecom) got together to design a better way to attach low-speed I/O devices to a computer. Since then, hundreds of other companies have joined them. The resulting standard, officially released in 1998, is called USB (Universal Serial Bus) and it is being widely implemented in personal computers. It is described further in Anderson (1997) and Tan (1997). Some of the goals of the companies that originally conceived of the USB and started the project were as follows:

218

THE DIGITAL LOGIC LEVEL

CHAP. 3

1. Users must not have to set switches or jumpers on boards or devices. 2. Users must not have to open the case to install new I/O devices. 3. There should be only one kind of cable, good for connecting all devices. 4. I/O devices should get their power from the cable. 5. Up to 127 devices should be attachable to a single computer. 6. The system should support real-time devices (e.g., sound, telephone). 7. Devices should be installable while the computer is running. 8. No reboot should be needed after installing a new device. 9. The new bus and its I/O devices should be inexpensive to manufacture. USB meets all these goals. It is designed for low-speed devices such as keyboards, mice, still cameras, snapshot scanners, digital telephones, and so on. Version 1.0 has a bandwidth of 1.5 Mbps, which is enough for keyboards and mice. Version 1.1 runs at 12 Mbps, which is enough for printers, digital cameras, and many other devices. These low limits were chosen to keep the cost down. A USB system consists of a root hub that plugs into the main bus (see Fig. 3-52). This hub has sockets for cables that can connect to I/O devices or to expansion hubs, to provide more sockets, so the topology of a USB system is a tree with its root at the root hub, inside the computer. The cables have different connectors on the hub end and on the device end, to prevent people from accidentally connecting two hub sockets together. The cable consists of four wires: two for data, one for power (+5 volts), and one for ground. The signaling system transmits a 0 as a voltage transition and a 1 as the absence of a voltage transition, so long runs of 0s generate a regular pulse stream. When a new I/O device is plugged in, the root hub detects this event and interrupts the operating system. The operating system then queries the device to find out what it is and how much USB bandwidth it needs. If the operating system decides that there is enough bandwidth for the device, it assigns the new device a unique address (1 – 127) and downloads this address and other information to configuration registers inside the device. In this way, new devices can be added on-the-fly, without any user configuration required and without having to install new ISA or PCI cards. Uninitialized cards start out with address 0, so they can be addressed. To make the cabling simpler, many USB devices contain built-in hubs to accept additional USB devices. For example, a monitor might have two hub sockets to accommodate the left and right speakers. Logically, the USB system can be viewed as a set of bit pipes from the root hub to the I/O devices. Each device can split its bit pipe up into at most 16 subpipes for different types of data (e.g., audio and video). Within each pipe or sub-

SEC. 3.6

219

EXAMPLE BUSES

pipe, data flows from the root hub to the device or the other way. There is no traffic between two I/O devices. Precisely every 1.00 ± 0.05 msec, the root hub broadcasts a new frame to keep all the devices synchronized in time. A frame is associated with a bit pipe, and consists of packets, the first of which is from the root hub to the device. Subsequent packets in the frame may also be in this direction, or they may be back from the device to the root hub. A sequence of four frames is shown in Fig. 3-59. Time (msec) 1

0

2

3

Idle Frame 1

Frame 0

Frame 2

Frame 3

Packets from root SOF

SOF

IN

DATA ACK

Packets from root SOF

SOF OUT DATA ACK From device

Data packet from device

SYN PID PAYLOAD CRC

SYN PID PAYLOAD CRC

Figure 3-59. The USB root hub sends out frames every 1.00 msec.

In Fig. 3-59 there is no work to be done in frames 0 and 2, so all that is needed is one SOF (Start of Frame) packet. This packet is always broadcast to all devices. Frame 1 is a poll, for example, a request to a scanner to return the bits it has found on the image it is scanning. Frame 3 consists of delivering data to some device, for example to a printer. USB supports four kinds of frames: control, isochronous, bulk, and interrupt. Control frames are used to configure devices, give them commands, and inquire about their status. Isochronous frames are for real-time devices such as microphones, loudspeakers, and telephones that need to send or accept data at precise time intervals. They have a highly-predictable delay but provide no retransmissions in the event of errors. Bulk frames are for large transfers to or from devices with no real-time requirements such as printers. Finally, interrupt frames are needed because USB does not support interrupts. For example, instead of having the keyboard cause an interrupt whenever a key is struck, the operating system can poll it every 50 msec to collect any pending keystrokes. A frame consists of one or more packets, possibly some in each direction. Four kinds of packets exist: token, data, handshake, and special. Token packets are from the root to a device and are for system control. The SOF, IN, and OUT packets in Fig. 3-59 are token packets. The SOF (Start of Frame) packet is the

220

THE DIGITAL LOGIC LEVEL

CHAP. 3

first one in each frame and marks the beginning of the frame. If there is no work to do, the SOF packet is the only one in the frame. The IN token packet is a poll, asking the device to return certain data. Fields in the IN packet tell which bit pipe is being polled so the device knows which data to return (if it has multiple streams). The OUT token packet announces that data for the device will follow. A fourth type of token packet, SETUP (not shown in the figure), is used for configuration. Besides the token packet, three other kinds exist. These are DATA (used to transmit up to 64 bytes of information either way), handshake, and special packets. The format of a data packet is shown in Fig. 3-59. It consists of an 8-bit synchronization field, an 8-bit packet type (PID), the payload, and a 16-bit CRC (Cyclic Redundancy Check) to detect errors. Three kinds of handshake packets are defined: ACK (the previous data packet was correctly received), NAK (a CRC error was detected), and STALL (please wait—I am busy right now). Now let us look at Fig. 3-59 again. Every 1.00 msec a frame must be sent from the root hub, even if there is no work. Frames 0 and 2 consist of just an SOF packet, indicating that there was no work. Frame 1 is a poll, so it starts out with SOF and IN packets from the computer to the I/O device, followed by a DATA packet from the device to the computer. The ACK packet tells the device that the data were received correctly. In case of an error, a NAK would be sent back to the device and the packet would be retransmitted for bulk data (but not for isochronous data). Frame 3 is similar in structure to frame 1, except that now the flow of data are from the computer to the device. After the USB standard was finalized in 1998, the people designing USB had nothing to do so they began working on a new high-speed version of USB, called USB 2.0. This standard is similar to the older USB 1.1 and backward compatible with it, except that it adds a third speed, 480 Mbps, to the two existing speeds. There are also some minor differences, such as the interface between the root hub and the controller. With USB 1.1 there were two interfaces available. The first one, UHCI (Universal Host Controller Interface), was designed by Intel and put most of the burden on the software designers (read: Microsoft). The second one, OHCI (Open Host Controller Interface), was designed by Microsoft and put most of the burden on the hardware designers (read: Intel). In USB 2.0 everyone agreed to a single new interface called EHCI (Enhanced Host Controller Interface). With USB now operating at 480 Mbps, it clearly competes with the IEEE 1394 serial bus popularly called FireWire, which runs at 400 Mbps. Although virtually every new Pentium system now comes with USB 2.0, 1394 is not likely to vanish because it has the backing of the consumer electronics industry. Camcorders, DVD players and similar devices will continue to be equipped with 1394 interfaces for the foreseeable future because the makers of these devices do not want to go to the expense of switching to a different standard that is hardly better than what they have now. Consumers, also, do not like changing standards.

SEC. 3.7

221

INTERFACING

3.7 INTERFACING A typical small- to medium-sized computer system consists of a CPU chip, memory chips, and some I/O controllers, all connected by a bus. We have already studied memories, CPUs, and buses in some detail. Now it is time to look at the last part of the puzzle, the I/O chips. It is through these chips that the computer communicates with the external world.

3.7.1 I/O Chips Numerous I/O chips are already available and new ones are being introduced all the time. Common chips include UARTs, USARTs, CRT controllers, disk controllers, and PIOs. A UART (Universal Asynchronous Receiver Transmitter) is a chip that can read a byte from the data bus and output it a bit at a time on a serial line for a terminal, or input data from a terminal. UARTs usually allow various speeds from 50 to 19,200 bps; character widths from 5 to 8 bits; 1, 1.5, or 2 stop bits; and provide even, odd, or no parity, all under program control. USARTs (Universal Synchronous Asynchronous Receiver Transmitters) can handle synchronous transmission using a variety of protocols as well as performing all the UART functions. Since we already looked at UARTs in Chap. 2, let us now study the parallel interface as an example of an I/O chip. PIO Chips A typical PIO (Parallel Input/Output) chip is the Intel 8255A, shown in Fig. 3-60. It has 24 I/O lines that can interface to any TTL-compatible device, for example, keyboards, switches, lights, or printers. In a nutshell, the CPU program can write a 0 or 1 to any line, or read the input status of any line, providing great flexibility. A small CPU-based system using a PIO can often replace a complete board full of SSI or MSI chips, especially in embedded systems. CS A0-A1

8255A Parallel I/O chip

WR RD RESET D0-D7

8

2

8

8

8

Port A

Port B

Port C

Figure 3-60. An 8255A PIO chip.

Although the CPU can configure the 8255A in many ways by loading status registers within the chip, we will concentrate on some of the simpler modes of

222

THE DIGITAL LOGIC LEVEL

CHAP. 3

operation. The simplest way of using the 8255A is as three independent 8-bit ports, A, B, and C. Associated with each port is an 8-bit latch register. To set the lines on a port, the CPU just writes an 8-bit number into the corresponding register, and the 8-bit number appears on the output lines and stays there until the register is rewritten. To use a port for input, the CPU just reads the corresponding register. Other operating modes provide for handshaking with external devices. For example, to output to a device that is not always ready to accept data, the 8255A can present data on an output port and wait for the device to send a pulse back saying that it has accepted the data and wants more. The necessary logic for latching such pulses and making them available to the CPU is included in the 8255A hardware. From the functional diagram of the 8255A we can see that in addition to 24 pins for the three ports, it has eight lines that connect directly to the data bus, a chip select line, read and write lines, two address lines, and a line for resetting the chip. The two address lines select one of the four internal registers, corresponding to ports A, B, C, and the status register, which has bits determining which ports are for input and which for output, and other functions. Normally, the two address lines are connected to the low-order bits of the address bus.

3.7.2 Address Decoding Up until now we have been deliberately vague about how chip select is asserted on the memory and I/O chips we have looked at. It is now time to look more carefully at how this is done. Let us consider a simple 16-bit embedded computer consisting of a CPU, a 2KB × 8 byte EPROM for the program, a 2KB × 8 byte RAM for the data, and a PIO. This small system might be used as a prototype for the brain of a cheap toy or simple appliance. Once in production, the EPROM might be replaced by a ROM. The PIO can be selected in one of two ways: as a true I/O device or as part of memory. If we choose to use it as an I/O device, then we must select it using an explicit bus line that indicates that an I/O device is being referenced, rather than memory. If we use the other approach, memory-mapped I/O, then we must assign it 4 bytes of the memory space for the three ports and the control register. The choice is somewhat arbitrary. We will choose memory-mapped I/O because it illustrates some interesting issues in I/O interfacing. The EPROM needs 2 KB of address space, the RAM also needs 2K of address space, and the PIO needs 4 bytes. Because our example address space is 64K, we must make a choice about where to put the three devices. One possible choice is shown in Fig. 3-61. The EPROM occupies addresses to 2K, the RAM occupies addresses 32 KB to 34 KB, and the PIO occupies the highest 4 bytes of the address space, 65532 to 65535. From the programmer’s point of view, it makes no difference which addresses are used; however, for interfacing it does matter. If

SEC. 3.7

223

INTERFACING

we had chosen to address the PIO via the I/O space, it would not need any memory addresses (but it would need four I/O space addresses). RAM at address 8000H

PIO at FFFCH

; ; ;; ;;;

EPROM at address 0

0

4K 8K 12K 16K 20K 24K 28K 32K 36K 40K 44K 48K 52K 56K 60K 64K

Figure 3-61. Location of the EPROM, RAM, and PIO in our 64 KB address space.

With the address assignments of Fig. 3-61, the EPROM should be selected by any 16-bit memory address of the form 00000xxxxxxxxxxx (binary). In other words, any address whose 5 high-order bits are all 0s falls in the bottom 2 KB of memory, hence in the EPROM. Thus the EPROM’s chip select could be wired to a 5-bit comparator, one of whose inputs was permanently wired to 00000. A better way to achieve the same effect is to use a five-input OR gate, with the five inputs attached to address lines!!!A11 to A15. If and only if all five lines are 0 will the output be 0, thus asserting CS (which is asserted low). Unfortunately, no five-input OR gate exists in the standard SSI series. The closest we can come is an eight-input NOR gate. By grounding three inputs and inverting the output we can nevertheless produce the correct signal, as shown in Fig. 3-61(a). SSI chips are so cheap that except in exceptional circumstances, using one inefficiently is not an issue. By convention, unused inputs are not shown in circuit diagrams. The same principle can be used for the RAM. However, the RAM should respond to binary addresses of the form 10000xxxxxxxxxxx, so an additional inverter is needed as shown in the figure. The PIO address decoding is somewhat more complicated, because it is selected by the four !!! addresses of the form 11111111111111xx. A possible circuit that asserts CS only when the correct address appears on the address bus is shown in the figure. It uses two eight-input NAND gates to feed an OR gate. To build the address decoding logic of Fig. 362(a) using SSI requires six chips—the four eight-input chips, an OR gate, and a chip with three inverters. However, if the computer really consists of only the CPU, two memory chips, and the PIO, we can use a trick to simplify greatly the address decoding. The trick is based on the fact that all EPROM addresses, and only EPROM addresses, !!! have a 0 in the high-order bit, A15. Therefore, we can just wire CS to A15 directly, as shown in Fig. 3-62(b). At this point the decision to put the RAM at 8000H may seem much less arbitrary. The RAM decoding can be done by noting that the only valid addresses of the form 10xxxxxxxxxxxxxx are in the RAM, so 2 bits of decoding are sufficient.

224

THE DIGITAL LOGIC LEVEL

CHAP. 3

A0 Address bus A15

CS

CS

2K 3 8 EPROM

2K 3 8 RAM

CS PI0

(a) A0 Address bus A15

CS

CS

2K 3 8 EPROM

2K 3 8 RAM

(b)

Figure 3-62. (a) Full address decoding. (b) Partial address decoding.

CS PI0

SEC. 3.7

INTERFACING

225

Similarly, any address starting with 11 must be a PIO address. The complete decoding logic is now two NAND gates and an inverter. Because an inverter can be made from a NAND gate by just tying the two inputs together, a single quad NAND chip is now more than sufficient. The address decoding logic of Fig. 3-62(b) is called partial address decoding, because the full addresses are not used. It has the property that a read from addresses 0001000000000000, 0001100000000000, or 0010000000000000 will give the same result. In fact, every address in the bottom half of the address space will select the EPROM. Because the extra addresses are not used, no harm is done, but if one is designing a computer that may be expanded in the future (an unlikely occurrence in a toy), partial decoding should be avoided because it ties up too much address space. Another common address decoding technique is to use a decoder, such as that shown in Fig. 3-13. By connecting the three inputs to the three high-order address lines, we get eight outputs, corresponding to addresses in the first 8K, second 8K, and so on. For a computer with eight RAMs, each 8K × 8, one such chip provides the complete decoding. For a computer with eight 2K × 8 memory chips, a single decoder is also sufficient, provided that the memory chips are each located in distinct 8-KB chunks of address space. (Remember our earlier remark that the position of the memory and I/O chips within the address space matters.)

3.8 SUMMARY Computers are constructed from integrated circuit chips containing tiny switching elements called gates. The most common gates are AND, OR, NAND, NOR, and NOT . Simple circuits can be built up by directly combining individual gates. More complex circuits are multiplexers, demultiplexers, encoders, decoders, shifters, and ALUs. Arbitrary Boolean functions can be programmed using a PLA. If many Boolean functions are needed, PLAs are often more efficient. The laws of Boolean algebra can be used to transform circuits from one form to another. In many cases more economical circuits can be produced this way. Computer arithmetic is done by adders. A single-bit full adder can be constructed from two half adders. An adder for a multibit word can be built by connecting multiple full adders in such a way as to allow the carry out of each one feed into its left-hand neighbor. The components of (static) memories are latches and flip-flops, each of which can store one bit of information. These can be combined linearly into octal latches and flip-flops or logarithmically into full-scale word-oriented memories. Memories are available as RAM, ROM, PROM, EPROM, EEPROM, and flash. Static RAMs need not be refreshed; they keep their stored values as long as the

226

THE DIGITAL LOGIC LEVEL

CHAP. 3

power remains on. Dynamic RAMs, on the other hand, must be refreshed periodically to compensate for leakage from the little capacitors on the chip. The components of a computer system are connected by buses. Many, but not all, of the pins on a typical CPU chip directly drive one bus line. The bus lines can be divided into address, data, and control lines. Synchronous buses are driven by a master clock. Asynchronous buses use full handshaking to synchronize the slave to the master. The Pentium 4 is an example of a modern CPU. Modern systems using it have a memory bus, a PCI bus, an ISA bus, and a USB bus. The PCI bus can transfer 64 bits at a time at a rate of 66 MHz, which makes it fast enough for nearly all peripherals, but not fast enough for memory. Switches, lights, printers, and many other I/O devices can be interfaced to computers using parallel I/O chips such as the 8255A. These chips can be configured to be part of the I/O space or the memory space, as needed. They can be fully decoded or partially decoded, depending on the application.

PROBLEMS 1. A logician drives into a drive-in restaurant and says, ‘‘I want a hamburger or a hot dog and french fries.’’ Unfortunately, the cook flunked out of sixth grade and does not know (or care) whether ‘‘and’’ has precedence over ‘‘or.’’ As far as he is concerned, one interpretation is as good as the other. Which of the following cases are valid interpretations of the order? (Note that in English ‘‘or’’ means ‘‘exclusive or.’’) a. Just a hamburger. b. Just a hot dog. c. Just french fries. d. A hot dog and french fries. e. A hamburger and french fries. f. A hot dog and a hamburger. g. All three. h. Nothing—the logician goes hungry for being a wiseguy. 2. A missionary lost in Southern California stops at a fork in the road. He knows that two motorcycle gangs inhabit the area, one of which always tells the truth and one of which always lies. He wants to know which road leads to Disneyland. What question should he ask? 3. Use a truth table to show that X = (X AND Y) OR (X AND NOT Y). 4. There exist four Boolean functions of a single variable and 16 functions of two variables. How many functions of three variables are there? Of n variables?

CHAP. 3

227

PROBLEMS

5. Show how the AND function can be constructed from two NAND gates. 6. Using the three-variable multiplexer chip of Fig. 3-12, implement a function whose output is the parity of the inputs, that is, the output is 1 if and only if an even number of inputs are 1. 7. Put on your thinking cap. The three-variable multiplexer chip of Fig. 3-12 is actually capable of computing an arbitrary function of four Boolean variables. Describe how, and as an example, draw the logic diagram for the function that is 0 if the English word for the truth table row has an even number of letters, 1 if it has an odd number of letters (e.g., 0000 = zero = four letters → 0; 0111 = seven = five letters → 1; 1101 = thirteen = eight letters → 0). Hint: If we call""the fourth input variable D, the eight input lines may be wired to Vcc , ground, D, or D. 8. Draw the logic diagram of a 2-bit encoder, a circuit with four input lines, exactly one of which is high at any instant, and two output lines whose 2-bit binary value tells which input is high. 9. Draw the logic diagram of a 2-bit demultiplexer, a circuit whose single input line is steered to one of the four output lines depending on the state of the two control lines. 10. Redraw the PLA of Fig. 3-15 in enough detail to show how the majority logic function of Fig. 3-3 can be implemented. In particular, be sure to show which connections are present in both matrices. 11. What does this circuit do?

A B

C D

12. A common MSI chip is a 4-bit adder. Four of these chips can be hooked up to form a 16-bit adder. How many pins would you expect the 4-bit adder chip to have? Why? 13. An n-bit adder can be constructed by cascading n full adders in series, with the carry into stage i, Ci , coming from the output of stage i − 1. The carry into stage 0, C 0 , is 0. If each stage takes T nsec to produce its sum and carry, the carry into stage i will not be valid until iT nsec after the start of the addition. For large n the time required for the carry to ripple through to the high-order stage may be unacceptably long. Design an adder that works faster. Hint: Each Ci can be expressed in terms of the operand bits Ai − 1 and Bi − 1 as well as the carry Ci − 1 . Using this relation it is possible to express Ci as a function of the inputs to stages 0 to i − 1, so all the carries can be generated simultaneously. 14. If all the gates in Fig. 3-19 have a propagation delay of 1 nsec, and all other delays can be ignored, what is the earliest time a circuit using this design can be sure of having a valid output bit?

228

THE DIGITAL LOGIC LEVEL

CHAP. 3

15. The ALU of Fig. 3-20 is capable of doing 8-bit 2’s complement additions. Is it also capable of doing 2’s complement subtractions? If so, explain how. If not, modify it to be able to do subtractions. 16. A 16-bit ALU is built up of 16 1-bit ALUs, each one having an add time of 10 nsec. If there is an additional 1 nsec delay for propagation from one ALU to the next, how long does it take for the result of a 16-bit add to appear? 17. Sometimes it is useful for an 8-bit ALU such as Fig. 3-20 to generate the constant −1 as output. Give two different ways this can be done. For each way, specify the values of the six control signals. 18. What is the quiescent state of the S and R inputs to an SR latch built of two gates?

NAND

19. The circuit of Fig. 3-26 is a flip-flop that is triggered on the rising edge of the clock. Modify this circuit to produce a flip-flop that is triggered on the falling edge of the clock. 20. The 4 × 3 memory of Fig. 3-29 uses 22 AND gates and three OR gates. If the circuit were to be expanded to 256 × 8, how many of each would be needed? 21. To help meet the payments on your new personal computer, you have taken up consulting for fledgling SSI chip manufacturers. One of your clients is thinking ""about putting out a chip containing four D flip-flops, each containing both Q and Q, on request of a potentially important customer. The proposed design has all four clock signals ganged together, also on request. Neither preset nor clear is present. Your assignment is to give a professional evaluation of the design. 22. As more and more memory is squeezed onto a single chip, the number of pins needed to address it also increases. It is often inconvenient to have large numbers of address pins on a chip. Devise a way to address 2n words of memory using fewer than n pins. 23. A computer with a 32-bit wide data bus uses 1M × 1 dynamic RAM memory chips. What is the smallest memory (in bytes) that this computer can have? 24. Referring to the timing diagram of Fig. 3-38, suppose that you slowed the clock down to a period of 20 nsec instead of 10 nsec as shown but the timing constraints remained unchanged. """""" How much time would the memory have to get the data onto the bus during T3 after MREQ was asserted, in the worst case? 25. Again referring to Fig. 3-38, suppose that the clock remained at 100 MHz, but TDS was increased to 4 nsec. Could 10-nsec memory chips be used? Can you envision a chip in 26. In Fig. 3-38(b), TML is specified to be at least 3 nsec. """""" which it is negative? Specifically, could the CPU assert MREQ before the address was stable? Why or why not? 27. Assume that the block transfer of Fig. 3-42 were done on the bus of Fig. 3-38. How much more bandwidth is obtained by using a block transfer over individual transfers for long blocks? Now assume that the bus is 32 bits wide instead of 8 bits wide. Answer the question again.

CHAP. 3

229

PROBLEMS

28. Denote the transition times of the address lines of Fig. 3-39 as TA1 and TA2 , and the """""" transition times of MREQ as TMREQ1 and TMREQ2 , and so on. Write down all the inequalities implied by the full handshake. 29. Most 32-bit buses permit 16-bit reads and writes. Is there any ambiguity about where to place the data? Discuss. 30. Many CPUs have a special bus cycle type for interrupt acknowledge. Why? 31. A 64-bit computer with a 200-MHz bus requires four cycles to read a 64-bit word. How much bus bandwidth does the CPU consume in the worst case? 32. A 32-bit CPU with address lines A2–A31 requires all memory references to be aligned. That is, words have to be addressed at multiples of 4 bytes, and half-words have to be addressed at even bytes. Bytes can be anywhere. How many legal combinations are there for memory reads, and how many pins are needed to express them? Give two answers and make a case for each one. 33. Why is it impossible for the Pentium 4 to work on a 32-bit PCI bus without losing any functionality? After all, other computers with a 64-bit data bus can do 32-bit, 16-bit, and even 8-bit wide transfers. 34. Suppose that a CPU has a level 1 cache and a level 2 cache, with access times of 1 nsec and 2 nsec, respectively. The main memory access time is 10 nsec. If 20% of the accesses are level 1 cache hits and 60% are level 2 cache hits, what is the average access time? 35. Is it likely that an 8051-based embedded system would include an 8255A chip? 36. Calculate the bus bandwidth needed to display a VGA (640 × 480) true-color movie at 30 frames/sec. Assume that the data must pass over the bus twice, once from the CDROM to the memory and once from the memory to the screen. 37. Which Pentium 4 signal do you think drives the PCI bus FRAME# line? 38. Which of the signals of Fig. 3-56 is not strictly necessary for the bus protocol to work? 39. A PCI Express system has 5 Mbps links (gross capacity). How many signal wires are needed in each direction for 8x operation? What is the gross capacity each way? What is the net capacity each way? 40. A computer has instructions that each require two bus cycles, one to fetch the instruction and one to fetch the data. Each bus cycle takes 10 nsec and each instruction takes 20 nsec (i.e., the internal processing time is negligible). The computer also has a disk with 2048 512-byte sectors per track. Disk rotation time is 5 msec. To what percent of its normal speed is the computer reduced during a DMA transfer if each 32-bit DMA transfer takes one bus cycle? 41. The maximum payload of an isochronous data packet on the USB bus is 1023 bytes. Assuming that a device may send only one data packet per frame, what is the maximum bandwidth for a single isochronous device? 42. What would the effect be of adding a third input line to the PIO of Fig. 3-62(b) if this new line were connected to A13?

NAND

gate selecting the

230

THE DIGITAL LOGIC LEVEL

CHAP. 3

43. Write a program to simulate the behavior of an m × n array of two-input NAND gates. This circuit, contained on a chip, has j input pins and k output pins. The values of j, k, m, and n are compile-time parameters of the simulation. The program should start off by reading in a ‘‘wiring list,’’ each wire of which specifies an input and an output. An input is either one of the j input pins or the output of some NAND gate. An output is either one of the k output pins or an input to some NAND gate. Unused inputs are logical 1. After reading in the wiring list, the program should print the output for each of the 2 j possible inputs. Gate array chips like this one are widely used for putting custom circuits on a chip because most of the work (depositing the gate array on the chip) is independent of the circuit to be implemented. Only the wiring is specific to each design. 44. Write a program to read in two arbitrary Boolean expressions and see if they represent the same function. The input language should include single letters, as Boolean variables, the operands AND, OR, and NOT, and parentheses. Each expression should fit on one input line. The program should compute the truth tables for both functions and compare them. 45. Write a program to read in a collection of Boolean expressions and compute the 24 × 50 and 50 × 6 matrices needed to implement them with the PLA of Fig. 3-15. The input language should be the same as the previous problem. Print the matrices on the line printer.

4 THE MICROARCHITECTURE LEVEL

The level above the digital logic level is the microarchitecture level. Its job is to implement the ISA (Instruction Set Architecture) level above it, as illustrated in Fig. 1-2. The design of the microarchitecture level depends on the ISA being implemented, as well as the cost and performance goals of the computer. Many modern ISAs, particularly RISC designs, have simple instructions that can usually be executed in a single clock cycle. More complex ISAs, such as the Pentium 4, may require many cycles to execute a single instruction. Executing an instruction may require locating the operands in memory, reading them, and storing results back into memory. The sequencing of operations within a single instruction often leads to a different approach to control than that for simple ISAs.

4.1 AN EXAMPLE MICROARCHITECTURE Ideally, we would like to introduce this subject by explaining the general principles of microarchitecture design. Unfortunately, there are no general principles; every one is a special case. Consequently, we will discuss a detailed example instead. For our example ISA, we have chosen a subset of the Java Virtual Machine, as we promised in Chap. 1. This subset contains only integer instructions, so we have named it IJVM. We will discuss the full JVM in Chap. 5. We will start out by describing the microarchitecture on top of which we will implement IJVM. IJVM has some complex instructions. Many such architectures 231

232

THE MICROARCHITECTURE LEVEL

CHAP. 4

have often been implemented through microprogramming, as discussed in Chap. 1. Although IJVM is small, it is a good starting point for describing the control and sequencing of instructions. Our microarchitecture will contain a microprogram (in ROM), whose job is to fetch, decode, and execute IJVM instructions. We cannot use the Sun JVM interpreter for the microprogram because we need a tiny microprogram that drives the individual gates in the actual hardware efficiently. In contrast, the Sun JVM interpreter was written in C for portability, and cannot control the hardware at the level of detail we need. Since the actual hardware used consists only of the basic components described in Chap. 3, in theory, after fully understanding this chapter, the reader should be able to go out and buy a large bag full of transistors and build this subset of the JVM machine. Students who successfully accomplish this task will be given extra credit (and a complete psychiatric examination). A convenient model for the design of the microarchitecture is to think of the design as a programming problem, where each instruction at the ISA level is a function to be called by a master program. In this model, the master program is a simple, endless loop that determines a function to be invoked, calls the function, then starts over, very much like Fig. 2-3. The microprogram has a set of variables, called the state of the computer, which can be accessed by all the functions. Each function changes at least some of the variables making up the state. For example, the Program Counter (PC) is part of the state. It indicates the memory location containing the next function (i.e., ISA instruction) to be executed. During the execution of each instruction, the PC is advanced to point to the next instruction to be executed. IJVM instructions are short and sweet. Each instruction has a few fields, usually one or two, each of which has some specific purpose. The first field of every instruction is the opcode (short for operation code), which identifies the instruction, telling whether it is an ADD or a BRANCH, or something else. Many instructions have an additional field, which specifies the operand. For example, instructions that access a local variable need a field to tell which variable. This model of execution, sometimes called the fetch-execute cycle, is useful in the abstract and may also be the basis for implementation for ISAs like IJVM that have complex instructions. Below we will describe how it works, what the microarchitecture looks like, and how it is controlled by the microinstructions, each of which controls the data path for one cycle. Together, the list of microinstructions forms the microprogram, which we will present and discuss in detail.

4.1.1 The Data Path The data path is that part of the CPU containing the ALU, its inputs, and its outputs. The data path of our example microarchitecture is shown in Fig. 4-1. While it has been carefully optimized for interpreting IJVM programs, it is fairly similar to the data path used in most machines. It contains a number of 32-bit

SEC. 4.1

AN EXAMPLE MICROARCHITECTURE

233

MAR

To and from main memory

Memory control registers

MDR

PC

MBR

SP

LV

Control signals Enable onto B bus

CPP

Write C bus to register TOS

OPC C bus

B bus

H A

ALU control

B

6

N Z

ALU

Shifter

2

Shifter control

Figure 4-1. The data path of the example microarchitecture used in this chapter.

registers, to which we have assigned symbolic names such as PC, SP, and MDR. Though some of these names are familiar, it is important to understand that these registers are accessible only at the microarchitecture level (by the microprogram). They are given these names because they usually hold a value corresponding to the variable of the same name in the ISA level architecture. Most registers can

234

THE MICROARCHITECTURE LEVEL

CHAP. 4

drive their contents onto the B bus. The output of the ALU drives the shifter and then the C bus, whose value can be written into one or more registers at the same time. There is no A bus for the moment; we will add one later. The ALU is identical to the one shown in Fig. 3-19 and Fig. 3-20. Its function is determined by six control lines. The short diagonal line labeled ‘‘6’’ in Fig. 4-1 indicates that there are six ALU control lines. These are F0 and F1 for determining the ALU operation, ENA and ENB for individually enabling the inputs, INVA for inverting the left input, and INC for forcing a carry into the low-order bit, effectively adding 1 to the result. However, not all 64 combinations of ALU control lines do something useful. Some of the more interesting combinations are shown in Fig. 4-2. Not all of these functions are needed for IJVM, but for the full JVM many of them would come in handy. In many cases, there are multiple possibilities for achieving the same result. In this table, + means arithmetic plus and − means arithmetic minus, so, for example −A means the two’s complement of A. ################################################# ! F ! F ! ENA ! ENB ! INVA ! INC ! Function ! 0 !################################################# ! 1 ! ! ! ! ! ! 0 ! 1 ! 1 0 !################################################# ! 0 ! ! 0 ! A ! ! 0 ! 1 ! 0 ! 1 ! ! 0 ! B ! 0 !################################################# ! ! ! ! ! ! " ! !################################################# ! 0 ! ! 0 ! A ! 0 ! 1 ! 1 1 " ! ! ! ! ! ! ! ! 1 ! 0 ! 1 0 !################################################# ! 1 ! ! 0 ! B ! ! 1 ! 1 ! 1 ! 1 ! ! 0 ! A+B ! 0 !################################################# ! ! ! ! ! ! ! 1 ! 1 ! 1 0 !################################################# ! 1 ! ! 1 ! A+B+1 ! ! 1 ! 1 ! 1 ! 0 ! ! 1 ! A+1 ! 0 !################################################# ! ! ! ! ! ! ! 0 1 ! 1 ! 0 !################################################# ! 1 ! ! 1 ! B+1 ! ! ! ! ! ! ! ! ! 1 1 1 1 1 1 B − A !################################################# ! ! ! ! ! ! ! !################################################# ! 1 ! ! 0 ! B−1 ! 1 ! 1 ! 0 1 ! ! ! ! ! ! ! ! 1 ! 1 ! 1 ! 1 ! 0 ! ! 1 ! −A ! ################################################# ! 0 ! 0 ! 1 ! 1 ! ! 0 ! A AND B ! 0 ################################################# ! ! ! ! ! ! ! ! 0 ! 0 ! 1 ! 1 ! 1 ! ! 0 ! A OR B ! ################################################# ! 0 ! 1 ! 0 ! 0 ! ! 0 ! 0 ! 0 ################################################# ! ! ! ! ! ! ! ! ! 1 ! 1 ! 0 ! 0 ! ! 1 ! 1 ! 0 ################################################# ! ! ! ! ! ! ! ! 1 1 0 0 1 0 −1 ! ################################################# ! ! ! ! ! ! ! Figure 4-2. Useful combinations of ALU signals and the function performed.

The ALU of Fig. 4-1 needs two data inputs: a left input (A) and a right input (B). Attached to the left input is a holding register, H. Attached to the right input is the B bus, which can be loaded from any one of nine sources, indicated by the nine gray arrows touching it. An alternative design, with two full buses, has a different set of trade-offs and will be discussed later in this chapter.

SEC. 4.1

AN EXAMPLE MICROARCHITECTURE

235

H can be loaded by choosing an ALU function that just passes the right input (from the B bus) through to the ALU output. One such function is adding the ALU inputs, only with ENA negated so the left input is forced to zero. Adding zero to the value on the B bus just yields the value on the B bus. This result can then be passed through the shifter unmodified and stored in H. In addition to the above functions, two other control lines can be used independently to control the output from the ALU. SLL8 (Shift Left Logical) shifts the contents left by 1 byte, filling the 8 least significant bits with zeros. SRA1 (Shift Right Arithmetic) shifts the contents right by 1 bit, leaving the most significant bit unchanged. It is explicitly possible to read and write the same register on one cycle. For example, it is allowed to put SP onto the B bus, disable the ALU’s left input, enable the INC signal, and store the result in SP, thus incrementing SP by 1 (see the eighth line in Fig. 4-2). How can a register be read and written on the same cycle without producing garbage? The solution is that reading and writing are actually performed at different times within the cycle. When a register is selected as the ALU’s right input, its value is put onto the B bus early in the cycle and kept there continuously throughout the entire cycle. The ALU then does its work, producing a result that passes through the shifter onto the C bus. Near the end of the cycle, when the ALU and shifter outputs are known to be stable, a clock signal triggers the store of the contents of the C bus into one or more of the registers. One of these registers may well be the one that supplied the B bus with its input. The precise timing of the data path makes it possible to read and write the same register on one cycle, as described below.

Data Path Timing The timing of these events is shown in Fig. 4-3. Here a short pulse is produced at the start of each clock cycle. It can be derived from the main clock, as shown in Fig. 3-21(c). On the falling edge of the pulse, the bits that will drive all the gates are set up. This takes a finite and known time, ∆w. Then the register needed on the B bus is selected and driven onto the B bus. It takes ∆x before the value is stable. Then the ALU and shifter begin to operate on valid data. After another ∆y, the ALU and shifter outputs are stable. After an additional ∆z, the results have propagated along the C bus to the registers, where they can be loaded on the rising edge of the next pulse. The load should be edge triggered and fast, so that even if some of the input registers are changed, the effects will not be felt on the C bus until long after the registers have been loaded. Also on the rising edge of the pulse, the register driving the B bus stops doing so, in preparation for the next cycle. MPC, MIR, and the memory are mentioned in the figure; their roles will be discussed shortly. It is important to realize that even though there are no storage elements in the data path, there is a finite propagation time through it. Changing the value on the

236

THE MICROARCHITECTURE LEVEL Registers loaded instantaneously from C bus and memory on rising edge of clock

Shifter output stable

Cycle 1 starts here

Clock cycle 1

∆w

CHAP. 4

∆x

Set up signals to drive data path Drive H and B bus

∆y

Clock cycle 2

New MPC used to load MIR with next microinstruction here

∆z

ALU and shifter

MPC available here

Propagation from shifter to registers

Figure 4-3. Timing diagram of one data path cycle.

B bus does not cause the C bus to change until a finite time later (due to the finite delays of each step). Consequently, even if a store changes one of the input registers, the value will be safely tucked away in the register long before the (nowincorrect) value being put on the B bus (or H) can reach the ALU. Making this design work requires rigid timing, a long clock cycle, a known minimum propagation time through the ALU, and a fast load of the registers from the C bus. However, with careful engineering, the data path can be designed so that it functions correctly all the time. Actual machines work this way. A somewhat different way to look at the data path cycle is to think of it as broken up into implicit subcycles. The start of subcycle 1 is triggered by the falling edge of the clock. The activities that go on during the subcycles are shown below along with the subcycle lengths (in parentheses). 1. The control signals are set up (∆w). 2. The registers are loaded onto the B bus (∆x). 3. The ALU and shifter operate (∆y). 4. The results propagate along the C bus back to the registers (∆z). At the rising edge of the next clock cycle, the results are stored in the registers. We said that the subcycles can be best thought of as being implicit. By this we mean there are no clock pulses or other explicit signals to tell the ALU when

SEC. 4.1

AN EXAMPLE MICROARCHITECTURE

237

to operate or tell the results to enter the C bus. In reality, the ALU and shifter run all the time. However, their inputs are garbage until a time ∆w + ∆x after the falling edge of the clock. Likewise, their outputs are garbage until ∆w + ∆x + ∆y has elapsed after the falling edge of the clock. The only explicit signals that drive the data path are the falling edge of the clock, which starts the data path cycle, and the rising edge of the clock, which loads the registers from the C bus. The other subcycle boundaries are implicitly determined by the inherent propagation times of the circuits involved. It is the responsibility of the design engineers to make sure that the time ∆w + ∆x + ∆y + ∆z comes sufficiently in advance of the rising edge of the clock to have the register loads work all the time. Memory Operation Our machine has two different ways to communicate with memory: a 32-bit, word-addressable memory port and an 8-bit, byte-addressable memory port. The 32-bit port is controlled by two registers, MAR (Memory Address Register), and MDR (Memory Data Register), as shown in Fig. 4-1. The 8-bit port is controlled by one register, PC, which reads 1 byte into the low-order 8 bits of MBR. This port can only read data from memory; it cannot write data to memory. Each of these registers (and all the other registers in Fig. 4-1) are driven by one or two control signals. An open arrow under a register indicates a control signal that enables the register’s output onto the B bus. Since MAR does not have a connection to the B bus, it does not have an enable signal. H does not have one either because it is always enabled, it being the only possible left ALU input. A solid black arrow under a register indicates a control signal that writes (i.e., loads) the register from the C bus. Since MBR cannot be loaded from the C bus, it does not have a write signal (although it does have two other enable signals, described below). To initiate a memory read or write, the appropriate memory registers must be loaded, then a read or write signal issued to the memory (not shown in Fig. 4-1). MAR contains word addresses, so that the values 0, 1, 2, etc., refer to consecutive words. PC contains byte addresses, so that the values 0, 1, 2, etc. refer to consecutive bytes. Thus putting a 2 in PC and starting a memory read will read out byte 2 from memory and put it in the low-order 8 bits of MBR. Putting a 2 in MAR and starting a memory read will read out bytes 8–11 (i.e., word 2) from memory and put them in MDR. This difference in functionality is needed because MAR and PC will be used to reference two different parts of memory. The need for this distinction will become clearer later. For the moment, suffice it to say that the MAR/MDR combination is used to read and write ISA-level data words and the PC/MBR combination is used to read the executable ISA-level program, which consists of a byte stream. All other registers that contain addresses use word addresses, like MAR.

238

THE MICROARCHITECTURE LEVEL

CHAP. 4

In the actual physical implementation, there is only one real memory and it is byte oriented. Allowing MAR to count in words (needed due to the way JVM is defined) while the physical memory counts in bytes is handled by a simple trick. When MAR is placed on the address bus, its 32 bits do not map onto the 32 address lines, 0 – 31, directly. Instead MAR bit 0 is wired to address bus line 2, MAR bit 1 is wired to address bus line 3, and so on. The upper 2 bits of MAR are discarded since they are only needed for word addresses above 232 , none of which are legal for our 4-GB machine. Using this mapping, when MAR is 1, address 4 is put onto the bus; when MAR is 2, address 8 is put onto the bus, and so forth. This trick is illustrated in Fig. 4-4. Discarded

32-Bit MAR (counts in words) 0 0

32-Bit address bus (counts in bytes)

Figure 4-4. Mapping of the bits in MAR to the address bus.

As mentioned above, data read from memory through the 8-bit memory port are returned in MBR, an 8-bit register. MBR can be gated (i.e., copied) onto the B bus in one of two ways: unsigned or signed. When the unsigned value is needed, the 32-bit word put onto the B bus contains the MBR value in the low-order 8 bits and zeros in the upper 24 bits. Unsigned values are useful for indexing into a table, or when a 16-bit integer has to be assembled from 2 consecutive (unsigned) bytes in the instruction stream. The other option for converting the 8-bit MBR to a 32-bit word is to treat it as a signed value between −128 and +127 and use this value to generate a 32-bit word with the same numerical value. This conversion is done by duplicating the MBR sign bit (leftmost bit) into the upper 24 bit positions of the B bus, a process known as sign extension. When this option is chosen, the upper 24 bits will either be all 0s or all 1s, depending on whether the leftmost bit of the 8-bit MBR is a 0 or a 1. The choice of whether the 8-bit MBR is converted to an unsigned or a signed 32-bit value on the B bus is determined by which of the two control signals (open arrows below MBR in Fig. 4-1) is asserted. The need for these two options is why two arrows are present. The ability to have the 8-bit MBR act like a 32-bit source to the B bus is indicated by the dashed box to the left of MBR in the figure.

SEC. 4.1

AN EXAMPLE MICROARCHITECTURE

239

4.1.2 Microinstructions To control the data path of Fig. 4-1, we need 29 signals. These can be divided into five functional groups, as described below. 9 Signals to control writing data from the C bus into registers. 9 Signals to control enabling registers onto the B bus for ALU input. 8 Signals to control the ALU and shifter functions. 2 Signals (not shown) to indicate memory read/write via MAR/MDR. 1 Signal (not shown) to indicate memory fetch via PC/MBR. The values of these 29 control signals specify the operations for one cycle of the data path. A cycle consists of gating values out of registers and onto the B bus, propagating the signals through the ALU and shifter, driving them onto the C bus, and finally writing the results in the appropriate register or registers. In addition, if a memory read data signal is asserted, the memory operation is started at the end of the data path cycle, after MAR has been loaded. The memory data are available at the very end of the following cycle in MBR or MDR and can be used in the cycle after that. In other words, a memory read on either port initiated at the end of cycle k delivers data that cannot be used in cycle k + 1, but only in cycle k + 2 or later. This seemingly counterintuitive behavior is explained by Fig. 4-3. The memory control signals are not generated in clock cycle 1 until just after MAR and PC are loaded at the rising edge of the clock, toward the end of clock cycle 1. We will assume the memory puts its results on the memory buses within one cycle so that MBR and/or MDR can be loaded on the next rising clock edge, along with the other registers. Put in other words, we load MAR at the end of a data path cycle and start the memory shortly thereafter. Consequently, we cannot really expect the results of a read operation to be in MDR at the start of the next cycle, especially if the width of the clock pulse is short. There is just not enough time if the memory takes one clock cycle. One data path cycle must intervene between starting a memory read and using the result. Of course, other operations can be performed during that cycle, just not ones that need the memory word. The assumption that the memory takes one cycle to operate is equivalent to assuming that the level 1 cache hit rate is 100%. This assumption is never true, but the complexity introduced by a variable-length memory cycle time is more than we want to deal with here. Since MBR and MDR are loaded on the rising edge of the clock, along with all the other registers, they may be read during cycles when a new memory read is being performed. They return the old values since the read has not yet had time to

240

THE MICROARCHITECTURE LEVEL

CHAP. 4

overwrite them. There is no ambiguity here; until new values are loaded into MBR and MDR at the rising edge of the clock, the previous values are still there and usable. Note that it is possible to perform back-to-back reads on two consecutive cycles since a read only takes 1 cycle. Also, both memories may operate at the same time. However, trying to read and write the same byte simultaneously gives undefined results. While it may be desirable to write the output on the C bus into more than one register, it is never desirable to enable more than one register onto the B bus at a time. (In fact, some real implementations will suffer physical damage if this is done.) With a small increase in circuitry, we can reduce the number of bits needed to select among the possible sources for driving the B bus. There are only nine possible input registers that can drive the B bus (where the signed and unsigned versions of MBR each count separately). Therefore, we can encode the B bus information in 4 bits and use a decoder to generate the 16 control signals, 7 of which are not needed. In a commercial design, the architects would experience an overwhelming urge to get rid of one of the registers so that 3 bits would do the job. As academics, we have the enormous luxury of being able to waste 1 bit to give a cleaner and simpler design. Bits

9

3

NEXT_ADDRESS

Addr

J M P C

J A M N

8 J A M Z

JAM

S L L 8

9

3

4

S F 0 F1 E E I I H O T C L S P M M W R F R P O P V P C D A R E E N N N N T I A R R T A C C S P A B V C 1 A E D H

ALU

C

B bus

B

Mem B bus registers 0 = MDR 1 = PC 2 = MBR 3 = MBRU 4 = SP

5 = LV 6 = CPP 7 = TOS 8 = OPC 9 -15 none

Figure 4-5. The microinstruction format for the Mic-1.

At this point we can control the data path with 9 + 4 + 8 + 2 + 1 = 24 signals, hence 24 bits. However, these 24 bits only control the data path for one cycle. The second part of the control is to determine what is to be done on the following cycle. To include this in the design of the controller, we will create a format for describing the operations to be performed using the 24 control bits plus two additional fields: the NEXT"ADDRESS field and the JAM field. The contents of each of these fields will be discussed shortly. Figure 4-5 shows a possible format, divided into the six groups, and containing the following 36 signals:

SEC. 4.1

AN EXAMPLE MICROARCHITECTURE

241

Addr – Contains the address of a potential next microinstruction. JAM

– Determines how the next microinstruction is selected.

ALU

– ALU and shifter functions.

C

– Selects which registers are written from the C bus.

Mem

– Memory functions.

B

– Selects the B bus source; it is encoded as shown.

The ordering of the groups is, in principle, arbitrary although we have actually chosen it very carefully to minimize line crossings in Fig. 4-6. Line crossings in schematic diagrams like Fig. 4-6 often correspond to wire crossings on chips, which cause trouble in two-dimensional designs and are best minimized.

4.1.3 Microinstruction Control: The Mic-1 So far we have described how the data path is controlled, but we have not yet described how it is decided which of the control signals should be enabled on each cycle. This is determined by a sequencer that is responsible for stepping through the sequence of operations necessary for the execution of a single ISA instruction. The sequencer must produce two kinds of information each cycle: 1. The state of every control signal in the system. 2. The address of the microinstruction that is to be executed next. Figure 4-6 is a detailed block diagram of the complete microarchitecture of our example machine, which we will call the Mic-1. It may look imposing initially but it is worth studying carefully. When you fully understand every box and every line in this figure, you will be well on your way to understanding the microarchitecture level. The block diagram has two parts: the data path, on the left, which we have already discussed in detail, and the control section, on the right, which we will now look at. The largest and most important item in the control portion of the machine is a memory called the control store. It is convenient to think of it as a memory that holds the complete microprogram, although it is sometimes implemented as a set of logic gates. In general, we will refer to it as the control store, to avoid confusion with the main memory, accessed through MBR and MDR. However, functionally, the control store is a memory that simply holds microinstructions instead of ISA instructions. For our example machine, it contains 512 words, each word consisting of one 36-bit microinstruction of the kind illustrated in Fig. 4-5. Actually, not all of these words are needed, but (for reasons to be explained shortly) we need addresses for 512 distinct words. In one important way, the control store is quite different from the main memory: instructions in main memory are always executed in address order (except for

242

THE MICROARCHITECTURE LEVEL

CHAP. 4

Memory control signals (rd, wr, fetch) 3

4 4-to-16 Decoder

MAR MDR

MPC

9

PC O

8

MBR SP

512 × 36-Bit control store for holding the microprogram

9

LV

JMPC

CPP

Addr

J

ALU

C

MIR M B

TOS JAMN/JAMZ

OPC H

B bus

ALU Shifter C bus

2 1-bit flip–flop

N

6 ALU control

High bit

Control signals Enable onto B bus

Z 2

Figure 4-6. The complete block diagram of our example microarchitecture, the Mic-1.

branches); microinstructions are not. The act of incrementing the program counter in Fig. 2-3 expresses the fact that the default instruction to execute after the current one is the instruction following the current one in memory. Microprograms need more flexibility (because microinstruction sequences tend to be short), so they usually do not have this property. Instead, each microinstruction explicitly specifies its successor.

Write C bus to register

SEC. 4.1

AN EXAMPLE MICROARCHITECTURE

243

Since the control store is functionally a (read-only) memory, it needs its own memory address register and its own memory data register. It does not need read and write signals, because it is continuously being read. We will call the control store’s memory address register MPC (MicroProgram Counter). This name is ironic since the locations in it are explicitly not ordered, so the concept of counting is not useful (but who are we to argue with tradition?). The memory data register is called MIR (MicroInstruction Register). Its function is to hold the current microinstruction, whose bits drive the control signals that operate the data path. The MIR register in Fig. 4-6 holds the same six groups as Fig. 4-5. The Addr and J (for JAM) groups control the selection of the next microinstruction and will be discussed shortly. The ALU group contains the 8 bits that select the ALU function and drive the shifter. The C bits cause individual registers to load the ALU output from the C bus. The M bits control memory operations. Finally, the last 4 bits drive the decoder that determines what goes onto the B bus. In this case we have chosen to use a standard 4-to-16 decoder, even though only nine possibilities are required. In a more finely-tuned design, a 4-to-9 decoder could be used. The trade-off here is using a standard circuit taken from a library of circuits versus designing a custom one. Using the standard circuit is simpler and is unlikely to introduce any bugs. Rolling your own uses less chip area but takes longer to design and you might get it wrong. The operation of Fig. 4-6 is as follows. At the start of each clock cycle (the falling edge of the clock in Fig. 4-3), MIR is loaded from the word in the control store pointed to by MPC. The MIR load time is indicated in the figure by ∆w. If one thinks in terms of subcycles, MIR is loaded during the first one. Once the microinstruction is set up in MIR, the various signals propagate out into the data path. A register is put out onto the B bus, the ALU knows which operation to perform, and there is lots of activity out there. This is the second subcycle. After an interval ∆w + ∆x from the start of the cycle, the ALU inputs are stable. Another ∆y later, everything has settled down and the ALU, N, Z, and shifter outputs are stable. The N and Z values are then saved in a pair of 1-bit flip-flops. These bits, like all the registers that are loaded from the C bus and from memory, are saved on the rising edge of the clock, near the end of the data path cycle. The ALU output is not latched but just fed into the shifter. The ALU and shifter activity occurs during subcycle 3. After an additional interval, ∆z, the shifter output has reached the registers via the C bus. Then the registers can be loaded near the end of the cycle (at the rising edge of the clock pulse in Fig. 4-3). Subcycle 4 consists of loading the registers and N and Z flip-flops. It terminates a little after the rising edge of the clock, when all the results have been saved and the results of the previous memory operations are available and MPC has been loaded. This process goes on and on until somebody gets bored with it and turns the machine off.

244

THE MICROARCHITECTURE LEVEL

CHAP. 4

In parallel with driving the data path, the microprogram has to determine which microinstruction to execute next, as they need not be run in the order they appear in the control store. The calculation of the address of the next microinstruction begins after MIR has been loaded and is stable. First, the 9-bit NEXT"ADDRESS field is copied to MPC. While this copy is taking place, the JAM field is inspected. If it has the value 000, nothing else is done; when the copy of NEXT"ADDRESS completes, MPC will point to the next microinstruction. If one or more of the JAM bits are 1, more work is needed. If JAMN is set, the 1-bit N flip-flop is ORed into the high-order bit of MPC. Similarly, if JAMZ is set, the 1-bit Z flip-flop is ORed there. If both are set, both are ORed there. The reason that the N and Z flip-flops are needed is that after the rising edge of the clock (while the clock is high), the B bus is no longer being driven, so the ALU outputs can no longer be assumed to be correct. Saving the ALU status flags in N and Z makes the correct values available and stable for the MPC computation, no matter what is going on around the ALU. In Fig. 4-6, the logic that does this computation is labeled ‘‘High bit.’’ The Boolean function it computes is F = (JAMZ AND Z) OR (JAMN AND N) OR NEXT"ADDRESS[8] Note that in all cases, MPC can take on only one of two possible values: 1. The value of NEXT"ADDRESS. 2. The value of NEXT"ADDRESS with the high-order bit ORed with 1. No other possibilities exist. If the high-order bit of NEXT"ADDRESS was already 1, using JAMN or JAMZ makes no sense. Note that when the JAM bits are all zeros, the address of the next microinstruction to be executed is simply the 9-bit number in its NEXT"ADDRESS field. When either JAMN or JAMZ are 1, there are two potential successors: NEXT"ADDRESS and NEXT"ADDRESS ORed with 0x100 (assuming that NEXT"ADDRESS ≤ 0xFF). (Note that 0x indicates that the number following it is in hexadecimal.) This point is illustrated in Fig. 4-7. The current microinstruction, at location 0x75, has NEXT"ADDRESS = 0x92 and JAMZ set to 1. Consequently, the next address of the microinstruction depends on the Z bit stored on the previous ALU operation. If the Z bit is 0, the next microinstruction comes from 0x92. If the Z bit is 1, the next microinstruction comes from 0x192. The third bit in the JAM field is JMPC. If it is set, the 8 MBR bits are bitwise ORed with the 8 low-order bits of the NEXT"ADDRESS field coming from the current microinstruction. The result is sent to MPC. The box with the label ‘‘O’’ in Fig. 4-6 does an OR of MBR with NEXT"ADDRESS if JMPC is 1 but just passes NEXT"ADDRESS through to MPC if JMPC is 0. When JMPC is 1, the low-order 8 bits of NEXT"ADDRESS are normally zero. The high-order bit can be 0 or 1, so the NEXT"ADDRESS value used with JMPC is normally 0x000 or 0x100. The reason for sometimes using 0x000 and sometimes using 0x100 will be discussed later.

SEC. 4.1

245

AN EXAMPLE MICROARCHITECTURE

Address

Addr

JAM

0x75

0x92

001

Data path control bits JAMZ bit set

…

0x92

…

0x192

One of these will follow 0x75 depending on Z

Figure 4-7. A microinstruction with JAMZ set to 1 has two potential successors.

The ability to OR MBR together with NEXT"ADDRESS and store the result in allows an efficient implementation of a multiway branch (jump). Notice that any of 256 addresses can be specified, determined solely by the bits present in MBR. In a typical use, MBR contains an opcode, so the use of JMPC will result in a unique selection for the next microinstruction to be executed for every possible opcode. This method is useful for quickly branching directly to the function corresponding to the just-fetched opcode. Understanding the timing of the machine is critical to what will follow, so it is perhaps worth repeating it again. We will do it in terms of subcycles, since this is easy to visualize, but the only real clock events are the falling edge, which starts the cycle, and the rising edge, which loads the registers and the N and Z flip-flops. Please refer to Fig. 4-3 once more. During subcycle 1, initiated by the falling edge of the clock, MIR is loaded from the address currently held in MPC. During subcycle 2, the signals from MIR propagate out and the B bus is loaded from the selected register. During subcycle 3, the ALU and shifter operate and produce a stable result. During subcycle 4, the C bus, memory buses, and ALU values become stable. At the rising edge of the clock, the registers are loaded from the C bus, N and Z flip-flops are loaded, and MBR and MDR get their results from the memory operation started at the end of the previous data path cycle (if any). As soon as MBR is available, MPC is loaded in preparation for the next microinstruction. Thus MPC gets its value sometime during the middle of the interval when the clock is high but after MBR/MDR are ready. It could be either level triggered (rather than edge triggered), or edge trigger a fixed delay after the rising edge of the clock. All that matters is that MPC is not loaded until the registers it depends on (MBR, N, and Z) are ready. As soon as the clock falls, MPC can address the control store and a new cycle can begin. Note that each cycle is self contained. It specifies what goes onto the B bus, what the ALU and shifter are to do, where the C bus is to be stored, and finally, what the next MPC value should be. One final note about Fig. 4-6 is worth making. We have been treating MPC as a proper register, with 9 bits of storage capacity that is loaded while the clock is MPC

246

THE MICROARCHITECTURE LEVEL

CHAP. 4

high. In reality, there is no need to have a register there at all. All of its inputs can be fed directly through to the control store. As long as they are present at the control store at the falling edge of the clock when MIR is selected and read out, that is sufficient. There is no need to actually store them in MPC. For this reason, MPC might well be implemented as a virtual register, which is just a gathering place for signals, more like an electronic patch panel, than a real register. Making MPC a virtual register simplifies the timing: now events only happen on the falling and rising edges of the clock and nowhere else. But if it is easier for you to think of MPC as a real register, that is also a valid viewpoint.

4.2 AN EXAMPLE ISA: IJVM Let us continue our example by introducing the ISA level of the machine to be interpreted by the microprogram running on the microarchitecture of Fig. 4-6 (IJVM). For convenience, we will sometimes refer to the Instruction Set Architecture as the macroarchitecture , to contrast it with the microarchitecture. Before we describe IJVM, however, we will digress slightly to motivate it.

4.2.1 Stacks Virtually all programming languages support the concept of procedures (methods), which have local variables. These variables can be accessed from inside the procedure but cease to be accessible once the procedure has returned. The question thus arises: ‘‘Where should these variables be kept in memory?’’ The simplest solution, to give each variable an absolute memory address, does not work. The problem is that a procedure may call itself. We will study these recursive procedures in Chap. 5. For the moment, suffice it to say that if a procedure is active (i.e., called) twice, it is impossible to store its variables in absolute memory locations because the second invocation will interfere with the first. Instead, a different strategy is used. An area of memory, called the stack, is reserved for variables, but individual variables do not get absolute addresses in it. Instead, a register, say, LV, is set to point to the base of the local variables for the current procedure. In Fig. 4-8(a), a procedure A, which has local variables a1, a2, and a3, has been called, so storage for its local variables has been reserved starting at the memory location pointed to by LV. Another register, SP, points to the highest word of A’s local variables. If LV is 100 and words are 4 bytes, then SP will be 108. Variables are referred to by giving their offset (distance) from LV. The data structure between LV and SP (and including both words pointed to) is called A’s local variable frame. Now let us consider what happens if A calls another procedure, B. Where should B’s four local variables (b1, b2, b3, b4) be stored? Answer: On the stack, on top of A’s, as shown in Fig. 4-8(b). Notice that LV has been adjusted by the

SEC. 4.2

SP

SP LV

247

AN EXAMPLE ISA: IJVM

a3 a2 a1 (a)

108 104 100

LV

b4 b3 b2 b1 a3 a2 a1 (b)

SP LV

c2 c1 b4 b3 b2 b1 a3 a2 a1

SP

LV

(c)

d5 d4 d3 d2 d1 a3 a2 a1 (d)

Figure 4-8. Use of a stack for storing local variables. (a) While A is active. (b) After A calls B. (c) After B calls C. (d) After C and B return and A calls D.

procedure call to point to B’s local variables instead of A’s. B’s local variables can be referred to by giving their offset from LV. Similarly, if B calls C, LV and SP are adjusted again to allocate space for C’s two variables, as shown in Fig. 4-8(c). When C returns, B becomes active again, and the stack is adjusted back to Fig. 4-8(b) so that LV now points to B’s local variables again. Likewise, when B returns, we get back to the situation of Fig. 4-8(a). Under all conditions, LV points to the base of the stack frame for the currently active procedure, and SP points to the top of the stack frame. Now suppose that A calls D, which has five local variables. We get the situation of Fig. 4-8(d), in which D’s local variables use the same memory that B’s did, as well as part of C’s. With this memory organization, memory is only allocated for procedures that are currently active. When a procedure returns, the memory used by its local variables is released. Stacks have another use, in addition to holding local variables. They can be used for holding operands during the computation of an arithmetic expression. When used this way, the stack is referred to as the operand stack. Suppose, for example, that before calling B, A has to do the computation a1 = a2 + a3;

One way of doing this sum is to push a2 onto the stack, as shown in Fig. 4-9(a). Here SP has been incremented by the number of bytes in a word, say, 4, and the first operand stored at the address now pointed to by SP. Next, a3 is pushed onto the stack, as shown in Fig. 4-9(b). As an aside on notation, we will typeset all program fragments in Helvetica , as above. We will also use this font for assembly language opcodes and machine registers, but in running text, program variables and procedures will be given in italics. The difference is that variables and procedure names are chosen by the user; opcodes and register names are built in. The actual computation can be done by now executing an instruction that pops two words off the stack, adds them together, and pushes the result back onto the stack, as shown in Fig. 4-9(c). Finally, the top word can be popped off the stack and stored back in local variable a1, as illustrated in Fig. 4-9(d).

;;; ;;; ;;; 248 SP LV

THE MICROARCHITECTURE LEVEL

a2 a3 a2 a1

(a)

SP

LV

a3 a2 a3 a2 a1

(b)

SP LV

a2 + a3 a3 a2 a1

CHAP. 4

SP LV

(c)

a3 a2 a2 + a3 (d)

Figure 4-9. Use of an operand stack for doing an arithmetic computation.

The local variable frames and the operand stacks can be intermixed. For example, when computing an expression like x 2 + f(x) part of the expression (e.g., x 2 ) may be on the operand stack when a function f is called. The result of the function is left on the stack, on top of x 2 , so the next instruction can add them. It is worth noting that while all machines use a stack for storing local variables, not all use an operand stack like this for doing arithmetic. In fact, most of them do not, but JVM and IJVM work like this, which is why we have introduced stack operations here. We will study them in more detail in Chap. 5.

4.2.2 The IJVM Memory Model We are now ready to look at the IJVM’s architecture. Basically, it consists of a memory that can be viewed in either of two ways: an array of 4,294,967,296 bytes (4 GB) or an array of 1,073,741,824 words, each consisting of 4 bytes. Unlike most ISAs, the Java Virtual Machine makes no absolute memory addresses directly visible at the ISA level, but there are several implicit addresses that provide the base for a pointer. IJVM instructions can only access memory by indexing from these pointers. At any time, the following areas of memory are defined: 1. The Constant Pool. This area cannot be written by an IJVM program and consists of constants, strings, and pointers to other areas of memory that can be referenced. It is loaded when the program is brought into memory and not changed afterward. There is an implicit register, CPP, that contains the address of the first word of the constant pool. 2. The Local Variable Frame. For each invocation of a method, an area is allocated for storing variables during the lifetime of the invocation. It is called the local variable frame. At the beginning of this frame reside the parameters (also called arguments) with which the method was invoked. The local variable frame does not include the operand stack, which is separate. However, for efficiency reasons, our implementation chooses to implement the operand stack immediately above the local variable frame. There is an implicit register that contains the address of the first location in the local variable frame. We will call this register LV. The parameters passed at the invocation of the method are stored at the beginning of the local variable frame.

SEC. 4.2

249

AN EXAMPLE ISA: IJVM

3. The Operand Stack. The stack frame is guaranteed not to exceed a certain size, computed in advance by the Java compiler. The operand stack space is allocated directly above the local variable frame, as illustrated in Fig. 4-10. In our implementation, it is convenient to think of the operand stack as part of the local variable frame. In any case, there is an implicit register that contains the address of the top word of the stack. Notice that, unlike CPP and LV, this pointer, SP, changes during the execution of the method as operands are pushed onto the stack or popped from it. 4. The Method Area. Finally, there is a region of memory containing the program, referred to as the ‘‘text’’ area in a UNIX process. There is an implicit register that contains the address of the instruction to be fetched next. This pointer is referred to as the Program Counter, or PC. Unlike the other regions of memory, the Method Area is treated as a byte array. Current Operand Stack 3

SP

Current Local Variable Frame 3 Local Variable Frame 2 Constant Pool

LV

Local Variable Frame 1

Method Area

CPP

Figure 4-10. The various parts of the IJVM memory.

One point needs to be made regarding the pointers. The CPP, LV, and SP registers are all pointers to words, not bytes, and are offset by the number of words. For the integer subset we have chosen, all references to items in the constant pool, the local variables frame, and the stack are words, and all offsets used to index into these frames are word offsets. For example, LV, LV + 1, and LV + 2 refer to the first three words of the local variables frame. In contrast, LV, LV + 4, and LV + 8 refer to words at intervals of four words (16 bytes). In contrast, PC contains a byte address, and an addition or subtraction to PC changes the address by a number of bytes, not a number of words. Addressing for

PC

250

THE MICROARCHITECTURE LEVEL

CHAP. 4

PC is different from the others, and this fact is apparent in the special memory port provided for PC on the Mic-1. Remember that it is only 1 byte wide. Incrementing PC by one and initiating a read results in a fetch of the next byte. Incrementing SP by one and initiating a read results in a fetch of the next word.

4.2.3 The IJVM Instruction Set The IJVM instruction set is shown in Fig. 4-11. Each instruction consists of an opcode and sometimes an operand, such as a memory offset or a constant. The first column gives the hexadecimal encoding of the instruction. The second gives its assembly language mnemonic. The third gives a brief description of its effect. ################################################################################# ! Hex ! ! ! Mnemonic Meaning ################################################################################# ! ! ! ! 0x10 ! BIPUSH byte !################################################################################# ! Push byte onto stack ! ! 0x59 ! DUP ! Copy top word on stack and push onto stack ! !################################################################################# ! ! ! !################################################################################# ! Unconditional branch ! 0xA7 ! GOTO offset ! ! ! ! 0x60 ! IADD !################################################################################# ! Pop two words from stack; push their sum ! !################################################################################# ! Pop two words from stack; push Boolean AND ! 0x7E ! IAND ! ! ! ! 0x99 ! IFEQ offset !################################################################################# ! Pop word from stack and branch if it is zero ! ! 0x9B ! IFLT offset ! Pop word from stack and branch if it is less than zero ! ################################################################################# ! ! ! ! 0x9F ! IF"ICMPEQ offset !################################################################################# ! Pop two words from stack; branch if equal ! ! 0x84 ! IINC varnum const ! Add a constant to a local variable ! !################################################################################# ! ! ! !################################################################################# ! Push local variable onto stack ! 0x15 ! ILOAD varnum ! ! ! ! 0xB6 ! INVOKEVIRTUAL disp ! Invoke a method !################################################################################# ! ! 0x80 ! IOR ! Pop two words from stack; push Boolean OR ! ################################################################################# ! ! ! ! 0xAC ! IRETURN !################################################################################# ! Return from method with integer value ! ! 0x36 ! ISTORE varnum ! Pop word from stack and store in local variable ! !################################################################################# ! ! ! 0x64 ! ISUB !################################################################################# ! Pop two words from stack; push their difference ! ! ! ! ! ################################################################################# ! 0x13 ! LDC"W index ! Push constant from constant pool onto stack ! ! 0x00 ! NOP ! Do nothing ! ################################################################################# ! ! ! ! ################################################################################# ! 0x57 ! POP ! Delete word on top of stack ! ! 0x5F ! SWAP ! Swap the two top words on the stack ! ################################################################################# ! ! ! ! ! 0xC4 ! WIDE ################################################################################# ! Prefix instruction; next instruction has a 16-bit index ! Figure 4-11. The IJVM instruction set. The operands byte, const, and varnum are 1 byte. The operands disp, index, and offset are 2 bytes.

Instructions are provided to push a word from various sources onto the stack. These sources include the constant pool (LDC"W), the local variable frame (ILOAD), and the instruction itself (BIPUSH). A variable can also be popped from

SEC. 4.2

AN EXAMPLE ISA: IJVM

251

the stack and stored into the local variable frame (ISTORE). Two arithmetic operations (IADD and ISUB) as well as two logical (Boolean) operations (IAND and IOR) can be performed using the two top words on the stack as operands. In all the arithmetic and logical operations, two words are popped from the stack and the result pushed back onto it. Four branch instructions are provided, one unconditional (GOTO) and three conditional ones (IFEQ, IFLT, and IF"ICMPEQ). All the branch instructions, if taken, adjust the value of PC by the size of their (16-bit signed) offset, which follows the opcode in the instruction. This offset is added to the address of the opcode. There are also IJVM instructions for swapping the top two words on the stack (SWAP), duplicating the top word (DUP), and removing it (POP). Some instructions have multiple formats, allowing a short form for commonly-used versions. In IJVM we have included two of the various mechanisms JVM uses to accomplish this. In one case we have skipped the short form in favor of the more general one. In another case we show how the prefix instruction WIDE can be used to modify the ensuing instruction. Finally, there is an instruction (INVOKEVIRTUAL) for invoking another method, and another instruction (IRETURN) for exiting the method and returning control to the method that invoked it. Due to the complexity of the mechanism we have slightly simplified the definition, making it possible to produce a straightforward mechanism for invoking a call and return. The restriction is that, unlike Java, we only allow a method to invoke a method existing within its own object. This restriction severely cripples the object orientation but allows us to present a much simpler mechanism, by avoiding the requirement to locate the method dynamically. (If you are not familiar with object-oriented programming, you can safely ignore this remark. What we have done is turn Java back into a nonobjectoriented language, such as C or Pascal.) On all computers except JVM, the address of the procedure to call is determined directly by the CALL instruction, so our approach is actually the normal case, not the exception. The mechanism for invoking a method is as follows. First, the caller pushes onto the stack a reference (pointer) to the object to be called. (This reference is not needed in IJVM since no other object may be specified, but it is retained for consistency with JVM.) In Fig. 4-12(a) this reference is indicated by OBJREF. Then the caller pushes the method’s parameters onto the stack, in this example, Parameter 1, Parameter 2, and Parameter 3. Finally, INVOKEVIRTUAL is executed. The INVOKEVIRTUAL instruction includes a displacement which indicates the position in the constant pool that contains the start address within the Method Area for the method being invoked. However, while the method code resides at the location pointed to by this pointer, the first 4 bytes in the method area contain special data. The first 2 bytes are interpreted as a 16-bit integer indicating the number of parameters for the method (the parameters themselves have previously been pushed onto the stack). For this count, OBJREF is counted as a parameter:

252

THE MICROARCHITECTURE LEVEL

CHAP. 4

parameter 0. This 16-bit integer, together with the value of SP, provides the location of OBJREF. Note that LV points to OBJREF rather than the first real parameter. The choice where LV points is somewhat arbitrary. The second 2 bytes in the method area are interpreted as another 16-bit integer indicating the size of the local variable area for the method being invoked. This is necessary because a new stack will be established for the method, beginning immediately above the local variable frame. Finally, the fifth byte in the method area contains the first opcode to be executed. Stack after INVOKEVIRTUAL Caller's LV Caller's PC Space for caller's local variables

Stack before INVOKEVIRTUAL Pushed parameters

Caller's local variable frame

Parameter 3 Parameter 2 Parameter 1 OBJREF Previous LV Previous PC

SP

Caller's local variables Parameter 2 Parameter 1 Link ptr (a)

SP

Stack base after INVOKEVIRTUAL

Stack base before INVOKEVIRTUAL LV

Parameter 3 Parameter 2 Parameter 1 Link ptr Previous LV Previous PC

LV

Caller's local variables Parameter 2 Parameter 1 Link ptr (b)

Figure 4-12. (a) Memory before executing INVOKEVIRTUAL. (b) After executing it.

The actual sequence that occurs for INVOKEVIRTUAL is as follows and is depicted in Fig. 4-12. The two unsigned index bytes that follow the opcode are used to construct an index into the constant pool table (the first byte is the high-order byte). The instruction computes the base address of the new local variable frame by subtracting off the number of parameters from the stack pointer and setting LV to point to OBJREF. At this location, overwriting OBJREF, the implementation stores the address of the location where the old PC is to be stored. This address is computed by adding the size of the local variable frame (parameters + local variables) to the address contained in LV. Immediately above the address where the old PC is to be stored is the address where the old LV is to be stored. Immediately above that address is the beginning of the stack for the newly-called procedure. SP is set to point to the old LV, which is the address immediately below the first empty location on the stack. Remember that SP always points to the top word on

SEC. 4.2

253

AN EXAMPLE ISA: IJVM

the stack. If the stack is empty, it points to the first location below the end of the stack because our stacks grow upward, toward higher addresses. In our figures, stacks always grow upward, toward the higher address at the top of the page. The last operation needed to carry out INVOKEVIRTUAL is to set PC to point to the fifth byte in the method code space. Stack before IRETURN Return value Previous LV Previous PC

SP

Caller's local variables Parameter 3 Parameter 2 Parameter 1 Link ptr Previous LV Previous PC Caller's local variable frame

Caller's local variables Parameter 2 Parameter 1 Link ptr (a)

Stack base before IRETURN LV

Stack after IRETURN Return value Previous LV Previous PC

Stack base after IRETURN

SP

Caller's local variables Parameter 2 Parameter 1 Link ptr

LV

(b)

Figure 4-13. (a) Memory before executing IRETURN. (b) After executing it.

The IRETURN instruction reverses the operations of the INVOKEVIRTUAL instruction, as shown in Fig. 4-13. It deallocates the space used by the returning method. It also restores the stack to its former state, except that (1) the (now overwritten) OBJREF word and all the parameters have been popped from the stack, and (2) the returned value has been placed at the top of the stack, at the location formerly occupied by OBJREF. To restore the old state, the IRETURN instruction must be able to restore the PC and LV pointers to their old values. It does this by accessing the link pointer (which is the word identified by the current LV pointer). In this location, remember, where the OBJREF was originally stored, the INVOKEVIRTUAL instruction stored the address containing the old PC. This word and the word above it are retrieved to restore PC and LV, respectively, to their old values. The return value, which is stored at the top of the stack of the terminating method, is copied to the location where the OBJREF was originally stored, and SP is restored to point to this location. Control is therefore returned to the instruction immediately following the INVOKEVIRTUAL instruction.

254

THE MICROARCHITECTURE LEVEL

CHAP. 4

So far, our machine does not have any input/output instructions. Nor are we going to add any. It does not need them any more than the Java Virtual Machine needs them, and the official specification for JVM never even mentions I/O. The theory is that a machine that does no input or output is ‘‘safe.’’ (Reading and writing are performed in JVM by means of calls to special I/O methods.

4.2.4 Compiling Java to IJVM Let us now see how Java and IJVM relate to one another. In Fig. 4-14(a) we show a simple fragment of Java code. When fed to a Java compiler, the compiler would probably produce the IJVM assembly language shown in Fig. 4-14(b). The line numbers from 1 to 15 at the left of the assembly language program are not part of the compiler output. Nor are the comments (starting with //). They are there to help explain a subsequent figure. The Java assembler would then translate the assembly program into the binary program shown in Fig. 4-14(c). (Actually, the Java compiler does its own assembly and produces the binary program directly.) For this example, we have assumed that i is local variable 1, j is local variable 2, and k is local variable 3. i = j + k; if (i == 3) k = 0; else j = j − 1;

(a)

1 ILOAD j 2 ILOAD k 3 IADD 4 ISTORE i 5 ILOAD i 6 BIPUSH 3 7 IF"ICMPEQ L1 8 ILOAD j 9 BIPUSH 1 10 ISUB 11 ISTORE j 12 GOTO L2 13 L1: BIPUSH 0 14 ISTORE k 15 L2: (b)

// i = j + k

// if (i == 3)

// j = j − 1

// k = 0

0x15 0x02 0x15 0x03 0x60 0x36 0x01 0x15 0x01 0x10 0x03 0x9F 0x00 0x0D 0x15 0x02 0x10 0x01 0x64 0x36 0x02 0xA7 0x00 0x07 0x10 0x00 0x36 0x03 (c)

Figure 4-14. (a) A Java fragment. (b) The corresponding Java assembly language. (c) The IJVM program in hexadecimal.

The compiled code is straightforward. First j and k are pushed onto the stack, added, and the result stored in i. Then i and the constant 3 are pushed onto the stack and compared. If they are equal, a branch is taken to L1, where k is set to 0. If they are unequal, the compare fails and code following IF"ICMPEQ is executed. When it is done, it branches to L2, where the then and else parts merge.

SEC. 4.2

255

AN EXAMPLE ISA: IJVM

The operand stack for the IJVM program of Fig. 4-14(b) is shown in Fig. 415. Before the code starts executing, the stack is empty, indicated by the horizontal line above the 0. After the first ILOAD, j is on the stack, as indicated by the boxed j above the 1 (meaning instruction 1 has executed). After the second ILOAD, two words are on the stack, as shown above the 2. After the IADD, there is only one word on the stack, and it contains the sum j + k. When the top word is popped from the stack and stored in i, the stack is empty, as shown above the 4.

0

j 1

k j 2

j+k 3

j 8

1 j 9

j–1 10

11

4

i 5

3 i 6

7

12

0 13

14

15

Figure 4-15. The stack after each instruction of Fig. 4-14(b).

Instruction 5 (ILOAD) starts the if statement by pushing i onto the stack (in 5) Next comes the constant 3 (in 6). After the comparison, the stack is empty again (7). Instruction 8 is the start of the else part of the Java program fragment. The else part continues until instruction 12, at which time it branches over the then part and goes to label L2.

4.3 AN EXAMPLE IMPLEMENTATION Having specified both the microarchitecture and the macroarchitecture in detail, the remaining issue is the implementation. In other words, what does a program running on the former and interpreting the latter look like, and how does it work? Before we can answer these questions, we must carefully consider the notation we will use to describe the implementation.

4.3.1 Microinstructions and Notation In principle, we could describe the control store in binary, 36 bits per word. But as in conventional programming languages, there is great benefit in introducing notation that conveys the essence of the issues we need to deal with while obscuring the details that can be ignored, or can be better handled automatically. It is important to realize here that the language we have chosen is intended to illustrate the concepts rather than to facilitate efficient designs. If the latter were our goal, we would use a different notation to maximize the flexibility available to the designer. One aspect where this issue is important is the choice of addresses.

256

THE MICROARCHITECTURE LEVEL

CHAP. 4

Since the memory is not logically ordered, there is no natural ‘‘next instruction’’ to be implied as we specify a sequence of operations. Much of the power of this control organization derives from the ability of the designer (or the assembler) to select addresses efficiently. We therefore begin by introducing a simple symbolic language that fully describes each operation without explaining fully how all addresses may have been determined. Our notation specifies all the activities that occur in a single clock cycle in a single line. We could, in theory, use a high-level language to describe the operations. However, cycle-by-cycle control is very important because it gives the opportunity to perform multiple operations concurrently, and it is necessary to be able to analyze each cycle to understand and verify the operations. If the goal is a fast, efficient implementation (other things being equal, fast and efficient is always better than slow and inefficient), then every cycle counts. In a real implementation, many subtle tricks are hidden in the program, using obscure sequences or operations in order to save a single cycle. There is a high payoff for saving cycles: a four-cycle instruction that can be reduced by two cycles now runs twice as fast. And this speedup is obtained every time we execute the instruction. One possible approach is simply to list the signals that should be activated each clock cycle. Suppose that in one cycle we want to increment the value of SP. We also want to initiate a read operation, and we want the next instruction to be the one residing at location 122 in the control store. We might write ReadRegister = SP, ALU = INC, WSP, Read, NEXT"ADDRESS = 122

where WSP means ‘‘write the SP register.’’ This notation is complete, but hard to understand. Instead we will combine the operations in a natural and intuitive way to capture the effect of what is happening: SP = SP + 1; rd

Let us call our high-level Micro Assembly Language ‘‘MAL’’ (French for ‘‘sick,’’ something you become if you have to write too much code in it). MAL is tailored to reflect the characteristics of the microarchitecture. During each cycle, any of the registers can be written, but typically only one is. Only one register can be gated to the B side of the ALU. On the A side, the choices are +1, 0, −1, and the register H. Thus we can use a simple assignment statement, as in Java, to indicate the operation to be performed. For example, to copy something from SP to MDR, we can say MDR = SP

To indicate the use of the ALU functions other than passing through the B bus, we can write, for example, MDR = H + SP

which adds the contents of the H register to SP and writes the result into MDR.

SEC. 4.3

AN EXAMPLE IMPLEMENTATION

257

The + operator is commutative (which means that the order of the operands does not matter), so the above statement can also be written as MDR = SP + H

and generate the same 36-bit microinstruction, even though strictly speaking H must be the left ALU operand. We have to be careful to use only legal operations. The most important legal operations are shown in the Fig. 4-16, where SOURCE can be any of MDR, PC, MBR, MBRU , SP, LV, CPP, TOS, or OPC (MBRU implies the unsigned version of MBR). These registers can all act as sources to the ALU on the B bus. Similarly, DEST can be any of MAR, MDR , PC, SP, LV, CPP, TOS, OPC, or H, all of which are possible destinations for the ALU output on the C bus. This format is deceptive because many seemingly reasonable statements are illegal. For example, MDR = SP + MDR

looks perfectly reasonable, but there is no way to execute it on the data path of Fig. 4-6 in one cycle. This restriction exists because for an addition (other than increment or decrement) one of the operands must be the H register. Likewise, H = H − MDR

might be useful, but it, too, is impossible, because the only possible source of a subtrahend (the value being subtracted) is the H register. It is up to the assembler to reject statements that look valid but are, in fact, illegal. We extend the notation to permit multiple assignments by the use of multiple equal signs. For example, adding 1 to SP and storing it back into SP as well as writing it into MDR can be accomplished by SP = MDR = SP + 1

To indicate memory reads and writes of 4-byte data words, we will just put rd and wr in the microinstruction. Fetching a byte through the 1-byte port is indicated by fetch. Assignments and memory operations can occur in the same cycle. This is indicated by writing them on the same line. To avoid any confusion, let us repeat the fact that the Mic-1 has two ways of accessing memory. Reads and writes of 4-byte data words use MAR/MDR and are indicated in the microinstructions by rd and wr, respectively. Reads of 1-byte opcodes from the instruction stream use PC/MBR and are indicated by fetch in the microinstructions. Both kinds of memory operations can proceed simultaneously. However, the same register may not receive a value from memory and the data path in the same cycle. Consider the code MAR = SP; rd MDR = H

The effect of the first microinstruction is to assign a value from memory to MDR at

258

THE MICROARCHITECTURE LEVEL

CHAP. 4

########################### ! DEST = H ! !###########################! DEST = SOURCE !########################### ! "" ! DEST = H ! !########################### ! """"""""" !########################### ! DEST = SOURCE ! ! DEST = H + SOURCE !###########################! ! DEST = H + SOURCE + 1 ! !###########################! DEST = H + 1 !########################### ! ! DEST = SOURCE + 1 ! !###########################! DEST = SOURCE − H !########################### ! ! ! DEST = SOURCE − 1 #! ##########################! ! DEST = −H #! ########################## ! ! DEST = H AND SOURCE ! #! ########################## ! DEST = H OR SOURCE ! #! ########################## ! DEST = 0 ! #! ########################## ! DEST = 1 ! #! ##########################! !!########################### !! DEST = −1 Figure 4-16. All permitted operations. Any of the above operations may be extended by adding ‘‘ 0 #! ################################################################################## ! ! ! ! Up ! Semaphore = semaphore + 1; ! Semaphore = semaphore + 1 ! ! ! if the other process was halted attempting to ! ! ! ! complete a down instruction on this sema! ! ! ! ! ! down and phore, it may now complete the ! ! ! ! ! continue running ! ! #! ################################################################################## ! ! ! ! ! Process halts until the other process ups this ! Semaphore = semaphore − 1 ! ! Down ! semaphore ! ! !################################################################################### Figure 6-27. The effect of a semaphore operation.

As mentioned above, Java has a language-based solution for dealing with race conditions, and we are discussing operating systems now. Thus we need a way to express semaphore usage in Java, even though it is not in the language or the standard classes. We will do this by assuming that two native methods have been written, up and down, which make the up and down system calls, respectively. By calling these with ordinary integers as parameters, we have a way to express the use of semaphores in Java programs. Figure 6-28 shows how the race condition can be eliminated through the use of semaphores. Two semaphores are added to the m class, available, which is initially 100 (the buffer size), and filled, which is initially 0. The producer starts executing at P1 in Fig. 6-28 and the consumer starts executing at C1 as before. The down call on filled halts the consumer processor immediately. When the producer has found the first prime, it calls down with available as parameter, setting available to 99. At P5 it calls up with filled as parameter, making filled 1. This action releases the consumer, which is now able to complete its suspended down call. At this point, filled is 0 and both processes are running. Let us now reexamine the race condition. At a certain point in time, in = 22, out = 21, the producer is at P1, and the consumer is at C5. The consumer finishes what it was doing and gets to C1 where it calls down on filled, which had the value 1 before the call and 0 after it. The consumer then takes the last number out of the buffer and ups available, making it 100. The consumer prints the number and goes to C1. Just before the consumer can call down, the producer finds the next prime and in quick succession executes statements P2, P3, and P4.

468

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

At this point, filled is 0. The producer is about to up it and the consumer is about to call down. If the consumer executes its instruction first, it will be suspended until the producer releases it (by calling up). On the other hand, if the producer goes first, the semaphore will be set to 1 and the consumer will not be suspended at all. In both cases, no wakeup is lost. This, of course, was our goal in introducing semaphores in the first place. The essential property of the semaphore operations is that they are indivisible. Once a semaphore operation has been initiated, no other process can use the semaphore until the first process has either completed the operation or been suspended trying. Furthermore, with semaphores, no wakeups are lost. In contrast, the if statements of Figure 6-25 are not indivisible. Between the evaluation of the condition and the execution of the selected statement, another process can send a wakeup signal. In effect the problem of process synchronization has been eliminated by declaring the up and down system calls made by up and down to be indivisible. In order for these operations to be indivisible, the operating system must prohibit two or more processes from using the same semaphore at the same time. At the very least, once an up or down system call has been made, no other user code will be run until the call has been completed. On single-processor systems, semaphores are sometimes implemented by disabling interrupts during semaphore operations. On multiple-processor systems, this trick does not work. Synchronization using semaphores is a technique that works for arbitrarily many processes. Several processes may be sleeping, attempting to complete a down system call on the same semaphore. When some other process finally performs an up on that semaphore, one of the waiting processes is allowed to complete its down and continue running. The semaphore value remains 0 and the other processes continue waiting. An analogy may make the nature of semaphores clearer. Imagine a picnic with 20 volleyball teams divided into 10 games (processes) each playing on its own court, and a large basket (the semaphore) for the volleyballs. Unfortunately, only seven volleyballs are available. At any instant, there are between zero and seven volleyballs in the basket (the semaphore has a value between 0 and 7). Putting a ball in the basket is an up because it increases the value of the semaphore; taking a ball out of the basket is a down because it decreases the value. At the start of the picnic, each court sends a player to the basket to get a volleyball. Seven of them successfully manage to get a volleyball (complete the down); three are forced to wait for a volleyball (i.e., fail to complete the down). Their games are suspended temporarily. Eventually, one of the other games finishes and puts a ball into the basket (executes an up). This operation allows one of the three players waiting around the basket to get a ball (complete an unfinished down), allowing one game to continue. The other two games remain suspended until two more balls are put into the basket. When two more balls come back (two more ups are executed), the last two games can proceed.

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

public class m { final public static int BUF"SIZE = 100; final public static long MAX"PRIME = 100000000000L; public static int in = 0, out = 0; public static long buffer[ ] = new long[BUF"SIZE]; public static producer p; public static consumer c; public static int filled = 0, available = 100; public static void main(String args[ ]) { p = new producer( ); c = new consumer( ); p.start( ); c.start( ); }

}

}

while (prime < m.MAX"PRIME) { prime = next"prime(prime); down(m.available); m.buffer[m.in] = prime; m.in = m.next(m.in); up(m.filled); }

private long next"prime(long prime){ ... }

class consumer extends Thread { native void up(int s); native void down(int s); public void run( ) { long emirp = 2;

}

// buffer runs from 0 to 99 // stop here // pointers to the data // primes stored here // name of the producer // name of the consumer // semaphores // main class // create the producer // create the consumer // start the producer // start the consumer

// This is a utility function for circularly incrementing in and out public static int next(int k) {if (k < BUF"SIZE − 1) return(k+1); else return(0);}

class producer extends Thread { native void up(int s); native void down(int s); public void run( ) { long prime = 2;

}

469

}

while (emirp < m.MAX"PRIME) { down(m.filled); emirp = m.buffer[m.out]; m.out = m.next(m.out); up(m.available); System.out.println(emirp); }

// producer class // methods on semaphores // producer code // scratch variable // statement P1 // statement P2 // statement P3 // statement P4 // statement P5

// function that computes next prime // consumer class // methods on semaphores // consumer code // scratch variable // statement C1 // statement C2 // statement C3 // statement C4 // statement C5

Figure 6-28. Parallel processing using semaphores.

470

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

6.4 EXAMPLE OPERATING SYSTEMS In this section we will continue discussing our example systems, the Pentium 4 and the UltraSPARC III. For each one we will look at an operating system used on that processor. For the Pentium 4 we will use Windows XP (called XP for short below); for the UltraSPARC III we will use UNIX. Since UNIX is simpler and in many ways more elegant, we will begin with it. Also, UNIX was designed and implemented first and had a major influence on XP, so this order makes more sense than the reverse.

6.4.1 Introduction In this section we will give a brief introduction to our two example operating systems, UNIX and XP, focusing on the history, structure, and system calls. UNIX UNIX was developed at Bell Labs in the early 1970s. The first version was written by Ken Thompson in assembler for the PDP-7 minicomputer. This was soon followed by a version for the PDP-11, written in a new language called C that was devised and implemented by Dennis Ritchie. In 1974, Ritchie and his colleague Ken Thompson published a landmark paper about UNIX (Ritchie and Thompson, 1974). For the work described in this paper they were later given the prestigious ACM Turing Award (Ritchie, 1984; Thompson, 1984). The publication of this paper stimulated many universities to ask Bell Labs for a copy of UNIX. Since Bell Labs’ parent company, AT&T, was a regulated monopoly at the time and was not permitted to be in the computer business, it had no objection to licensing UNIX to universities for a modest fee. In one of those coincidences that often shape history, the PDP-11 was the computer of choice at nearly all university computer science departments, and the operating systems that came with the PDP-11 were widely regarded as being dreadful by professors and students alike. UNIX quickly filled the void, not in the least because it was supplied with the complete source code, so people could, and did, tinker with it endlessly. One of the many universities that acquired UNIX early on was the University of California at Berkeley. Because the complete source code was available, Berkeley was able to modify the system substantially. Foremost among the changes was a port to the VAX minicomputer and the addition of paged virtual memory, the extension of file names from 14 characters to 255 characters, and the inclusion of the TCP/IP networking protocol, which is now used on the Internet (largely due to the fact that it was in Berkeley UNIX).

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

471

While Berkeley was making all these changes, AT&T itself continued to develop UNIX, leading to System III in 1982 and then System V in 1984. By the late 1980s, two different, and quite incompatible, versions of UNIX were in widespread use: Berkeley UNIX and System V. This split in the UNIX world, together with the fact that there were no standards for binary program formats, greatly inhibited the commercial success of UNIX because it was impossible for software vendors to write and package UNIX programs with the expectation that they would run on any UNIX system (as was routinely done with MS-DOS). After much bickering, a standard called POSIX (Portable Operating System-IX) was created by the IEEE Standards Board. is also known by its IEEE Standards number, P1003. It later became an International Standard. The standard is divided into many parts, each covering a different area of UNIX. The first part, P1003.1, defines the system calls; the second part, P1003.2, defines the basic utility programs, and so on. The P1003.1 standard defines about 60 system calls that all conformant systems must support. These are the basic calls for reading and writing files, creating new processes, and so on. Nearly all UNIX systems now support the P1003.1 system calls. However many UNIX systems also support extra system calls, especially those defined by System V and/or those in Berkeley UNIX. Typically these add up to 200 system calls to the set. In 1987, the author released the source code for a small version of UNIX, called MINIX, for use at universities (Tanenbaum, 1987). One of the students who studied MINIX at his university in Helsinki and ran it on his home PC was Linus Torvalds. After becoming thoroughly familiar with MINIX, Torvalds decided to write his own clone of MINIX, which was called Linux and has become quite popular. Both MINIX and Linux are conformant, and nearly everything said about UNIX in this chapter also applies to them unless stated otherwise. Although all these UNIX variants different greatly inside, for the most part in this chapter we will be discussing their system call interface, which is a superset of in all cases. The operating system for the UltraSPARC III is based on System V and is called Solaris. It also supports many of the Berkeley system calls. A rough breakdown of the Solaris system calls by category is given in Fig. 629. The file and directory management system calls are largest and the most important categories. Most of these come from P1003.1. A relatively large fraction of the others are derived from System V. One area that is largely due to Berkeley UNIX rather than System V is networking. Berkeley invented the concept of a socket, which is the endpoint of a network connection. The four-pin wall plugs to which telephones can be connected served as the model for this concept. It is possible for a UNIX process to create a socket, attach to it, and establish a connection to a socket on a distant machine. Over this connection it can then exchange data in both directions, typically using the TCP/IP protocol. Since networking technology has been in UNIX for decades and is very stable and mature, a substantial fraction of the servers on the Internet run UNIX.

472

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

#################################################################################### ! ! ! Category Some examples #################################################################################### ! ! ! ! File management ! Open, read, write, close, and lock files ! #################################################################################### ! Directory management ! Create and delete directories; move files around ! #################################################################################### ! ! ! ! Process management ! Spawn, terminate, trace, and signal processes ! #################################################################################### ! ! ! Memory management Share memory among processes; protect pages #################################################################################### ! ! ! ! Getting/setting parameters ! Get user, group, process ID; set priority ! #################################################################################### ! ! ! ! Dates and times ! Set file access times; use interval timer; profile execution ! #################################################################################### ! Networking ! Establish/accept connection; send/receive message ! #################################################################################### ! ! ! ! Miscellaneous #################################################################################### ! Enable accounting; manipulate disk quotas; reboot the system ! Figure 6-29. A rough breakdown of the UNIX system calls.

Since there are many implementations of UNIX, it is difficult to say much about the structure of the operating system since each one is somewhat different from all the others. However, in general, Fig. 6-30 applies to most of them. At the bottom, there is a layer of device drivers that shield the file system from the bare hardware. Originally, each device driver was written as an independent entity, separate from all the others. This arrangement led to a lot of duplicated effort, since many drivers must deal with flow control, error handling, priorities, separating data from control, and so on. This observation led Dennis Ritchie to develop a framework called streams for writing drivers in a modular way. With a stream, it is possible to establish a two-way connection from a user process to a hardware device and to insert one or more modules along the path. The user process pushes data into the stream, which then is processed or transformed by each module until it gets to the hardware. The inverse processing occurs for incoming data. Shell

User program

User mode

System call interface File system

Process management

Block cache

IPC

Scheduling

Device drivers

Signals

Memory mgmt.

Hardware

Figure 6-30. The structure of a typical UNIX system.

Kernel mode

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

473

On top of the device drivers comes the file system. It manages file names, directories, disk block allocation, protection, and much more. Part of the file system is a block cache, for holding the blocks most recently read in from disk, in case they are needed again soon. A variety of file systems have been used over the years, including the Berkeley fast file system (McKusick et al., 1984), and log-structured file systems (Rosenblum and Ousterhout, 1991; and Seltzer et al., 1993). The other part of the UNIX kernel is the process management portion. Among its various other functions, it handles IPC (InterProcess Communication), which allows processes to communicate with one another and synchronize to avoid race conditions. A variety of mechanisms are provided. The process management code also handles process scheduling, which is based on priorities. Signals, which are a form of (asynchronous) software interrupt, are also managed here. Finally, memory management is done here as well. Most UNIX systems support demandpaged virtual memory, sometimes with a few extra features, such as the ability of multiple processes to share common regions of address space. From its inception, UNIX has tried to be a small system, in order to enhance reliability and performance. The first versions of UNIX were entirely text-based, using terminals that could display 24 or 25 lines of 80 ASCII characters. The user interface was handled by a user-level program called the shell, which offered a command line interface. Since the shell was not part of the kernel, adding new shells to UNIX was easy, and over time a number of increasingly sophisticated ones were invented. Later on, when graphics terminals came into existence, a windowing system for UNIX, called X Windows, was developed at M.I.T. Still later, a full-fledged GUI (Graphical User Interface), called Motif, was put on top of X Windows. In keeping with the UNIX philosophy of having a small kernel, nearly all the code of X Windows and Motif runs in user mode, outside the kernel. Windows XP When the original IBM PC was launched in 1981, it came equipped with a 16-bit real-mode, single-user, command-line oriented operating system called MS-DOS 1.0 This operating system consisted of 8 KB of memory resident code. Two years later, a much more powerful 24-KB system, MS-DOS 2.0, appeared. It contained a command line processor (shell), with a number of features borrowed from UNIX. When IBM released the 286-based PC/AT in 1984, it came equipped with MS-DOS 3.0, by now 36 KB. Over the years, MS-DOS continued to acquire new features, but it was still a command-line oriented system. Inspired by the success of the Apple Macintosh, Microsoft decided to give MS-DOS a graphical user interface that it called Windows. The first three versions of Windows, culminating in Windows 3.x, were not true operating systems, but graphical user interfaces on top of MS-DOS, which was still in control of the

474

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

machine. All programs ran in the same address space and a bug in any one of them could bring the whole system to a grinding halt. The release of Windows 95 in 1995 still did not eliminate MS-DOS, although it introduced a new version, 7.0. Together, Windows 95 and MS-DOS 7.0 contained most of the features of a full-blown operating system, including virtual memory, process management, and multiprogramming. However, Windows 95 was not a full 32-bit program. It contained large chunks of old 16-bit code (as well as some 32-bit code) and still used the MS-DOS file system, with nearly all its limitations. The only major changes to the file system were the addition of long file names in place of the 8 + 3 character file names allowed in MS-DOS and the ability to have more than 65,536 blocks on a disk. Even with the release of Windows 98 in 1998, MS-DOS was still there (now called version 7.1) and running 16-bit code. Although a bit more functionality migrated from the MS-DOS part to the Windows part, and a disk layout suitable for larger disks was now standard, under the hood, Windows 98 was not very different from Windows 95. The main difference was the user interface, which integrated the desktop, the Internet, and television more closely. It was precisely this integration that attracted the attention of the U.S. Dept. of Justice, which then sued Microsoft claiming that it was an illegal monopoly. Windows 98 was followed by the short-lived Windows Millennium Edition (ME), which was a slightly improved Windows 98. While all these developments were going on, Microsoft was also busy with a completely new 32-bit operating system being written from the ground up. This new system was called Windows New Technology, or Windows NT. It was initially hyped as the replacement for all other operating systems for Intel-based PCs, but it was somewhat slow to catch on and was later redirected to the upper end of the market, where it found a niche. The second version of NT was called Windows 2000 and became the mainstream version, also for the desktop market. The successor to Windows 2000 was XP, but the changes here were relatively minor. XP is essentially a slightly improved Windows 2000. XP is sold in two versions: server and client. These two versions are nearly identical and are generated from the same source code. The server version is intended for machines that run as LAN-based file and print servers and has more elaborate management features than the client version, which is intended for desktop computing for a single user. The server version has a variant (enterprise) intended for large sites. The various versions are tuned differently, each one optimized for its expected environment. Other than these minor differences, all the versions are essentially the same. In fact, nearly all the executable files are identical for all versions. XP itself discovers which version it is by looking at a variable in an internal data structure (the registry). Users are forbidden by the license from changing this variable and thus converting the (inexpensive) client version into the (much more expensive) server or enterprise versions. We will not make any further distinction between these versions.

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

475

MS-DOS and all previous versions of Windows were single-user systems. XP, however, supports multiprogramming, so several users can work on the same machine at the same time. For example, a network server may have multiple users logged in simultaneously over a network, each accessing its own files. XP is a true 32-bit multiprogramming operating system. It supports multiple user processes, each of which has a full 32-bit demand-paged virtual address space. In addition, the system itself is written as 32-bit code everywhere. One of NT’s original improvements over Windows 95 was its modular structure. It consisted of a moderately small kernel that ran in kernel mode, plus a number of server processes that ran in user mode. User processes interacted with the server processes using the client-server model: a client sent a request message to a server, and the server did the work and returned the result to the client via a second message. This modular structure made it easier to port it to several computers besides the Intel line, including the DEC Alpha, IBM PowerPC, and SGI MIPS. However, for performance reasons, starting with NT 4.0, most of the system was put back into the kernel. Although we could go on for a long time both about how XP is structured internally and what its system call interface is like. Since our primary interest here is the virtual machine presented by various operating systems (i.e., the system calls), we will give a brief summary of the system structure and then move on to the system call interface. The structure of XP is illustrated in Fig. 6-31. It consists of a number of modules that are structured in layers and work together to implement the operating system. Each module has some particular function and a well-defined interface to the other modules. Nearly all the modules are written in C, although part of the graphics device interface is written in C++ and a tiny bit of the lowest layers are written in assembly language. At the bottom is a thin layer called the hardware abstraction layer. Its job is to present the rest of the operating system with abstract hardware devices, devoid of the warts and idiosyncracies with which real hardware is so richly endowed. Among the devices modeled are off-chip caches, timers, I/O buses, interrupt controllers, and DMA controllers. By exposing these to the rest of the operating system in idealized form, it becomes easier to port XP to other hardware platforms, since most of the modifications required are concentrated in one place. Above the hardware abstraction layer is a layer containing the kernel and the device drivers. The kernel and all the device drivers have direct access to the hardware when needed, as they contain hardware-dependent code. The kernel supports the primitive kernel objects, interrupt, trap, and exception handling, process scheduling and synchronization, multiprocessor synchronization, and time management. The purpose of this layer is to make the rest of the operating system completely independent of the hardware, and thus highly portable. The kernel is permanently resident in main memory and is not preemptible, although it can temporarily give up control to service I/O interrupts.

476

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

Win32 program User mode

Win32 subsystem System interface System services Executive

I/O File systems

File cache

Virtual memory

Processes and threads

Security

and

Object management Device drivers

Win32

Kernel

Graphics device interface

Kernel mode

Hardware abstraction layer Hardware

Figure 6-31. The structure of Windows XP.

Each device driver can control one or more I/O devices, but a device driver can also do things not related to a specific device, such as encrypting a data stream or even just providing access to kernel data structures. Since users can install new device drivers, they have the power to affect the kernel and corrupt the system. For this reason, drivers must be written with great care. Above the kernel and device drivers is the upper portion of the operating system, called the executive. The executive is architecture independent and can be ported to new machines with only a relatively small amount of effort. It consists of three layers. The lowest layer contains the file systems and the object manager. The file systems support the use of files and directories The object manager handles objects known to the kernel. These include processes, threads (lightweight processes within an address space), files, directories, semaphores, I/O devices, timers, and many others. The object manager also manages a namespace in which newly created objects can be placed so they can be referred to later. The next layer consists of six major parts, as shown in Fig. 6-31. The I/O manager provides a framework for managing I/O devices and provides generic I/O services. It uses the services of the file system, which in turn uses the device drivers, as well as the services of the object manager. The file cache manager is involved with managing file blocks and helping the virtual memory manager to determine which ones to keep in memory for future use. It is also involved with managing files that are mapped onto memory.

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

477

It is possible to configure XP with multiple file systems, in which case the cache manager works for all of them. When a block is needed, the cache manager is asked to supply it. If it does not have the block, the cache manager calls upon the appropriate file system to get it. Since files can be mapped into processes’ address spaces, the cache manager must interact with the virtual memory manager to provide the necessary consistency. The virtual memory manager implements XP’s demand-paged virtual memory architecture. It manages the mapping of virtual pages onto physical page frames. It thereby enforces the protection rules that restrict each process to only access those pages belonging to its address space and not to other processes’ address spaces (except under special circumstances). It also handles certain system calls that relate to virtual memory. The process and thread manager handles processes and threads, including their creation and destruction. It is concerned about the mechanisms used to manage them, rather than policies about how they are used. The security reference monitor enforces XP’s elaborate security mechanism, which meets the U.S. Dept. of Defense’s Orange Book C2 requirements. The Orange Book specifies a large number of rules that a conforming system must meet, starting with authenticated login through how access control is handled to the fact that virtual pages must be zeroed out before being reused. The graphics device interface handles image management for the monitor and printers. It provides system calls to allow user programs to write on the monitor or printers in a device-independent way. It also contains the hardware device drivers for graphical output. In versions of XP prior to XP 4.0, it was in user space but the performance was disappointing, so Microsoft moved it into the kernel to speed it up. The Win32 module also handles most of the system calls. It, too, was originally in user space but was also moved to the kernel to improve performance. On the top of the executive is a thin layer called system services. Its function is to provide an interface to the executive. It accepts the true XP system calls and calls other parts of the executive to have them executed. Outside the kernel are the user programs and the environment subsystem. The environmental subsystem is provided because user programs are not encouraged to make system calls directly (although they are technically capable of it). Instead, the environmental subsystem exports a set of function calls that user programs can use. Originally there were three environmental subsystems, Win32 (for NT, Windows 2000, XP or even Windows 95/98/ME programs), POSIX (for UNIX programs that have been ported), and OS/2 (for OS/2 programs that have been ported). Of these, only Win32 is supported. However, a new Services for UNIX module now exists and provides modest UNIX support. Windows applications use the Win32 functions and communicate with the Win32 subsystem to make system calls. The Win32 subsystem accepts the Win32 function calls (see below) and uses the system interface library module

478

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

(actually, a DLL file—see Chap. 7) to make the necessary true XP system calls to carry them out. This is how system calls are made in XP. Having looked briefly at the structure of XP, let us now turn to our main subject, the services offered by XP. This interface is the programmer’s main connection to the system. Unfortunately, Microsoft has never made the complete list of XP system calls public, and it also changes them from release to release. Under such conditions, writing programs that make system calls directly is nearly impossible. What Microsoft did do was define a set of calls called the Win32 API (Application Programming Interface) that are publicly known. These are library procedures that either make system calls to get the work done, or, in some case, do the work right in the user-space library procedure or in the Win32 subsystem. The Win32 API calls do not change with new releases, to promote stability. However, there are also XP API calls that may change between releases of XP. Although the Win32 API calls are not all XP system calls, it is better to focus on these here rather than the true XP system calls because the Win32 API calls are well documented and more stable over time. When Windows was ported to 64-bit machines, Microsoft changed the name of Win32 to cover both the 32-bit and 64bit versions, but for our purposes, looking at the 32-bit version is sufficient. The Win32 API philosophy is completely different from the UNIX philosophy. In the latter, the system calls are all publicly known and form a minimal interface: removing even one of them would reduce the functionality of the operating system. The Win32 philosophy is to provide a very comprehensive interface, often with three or four ways of doing the same thing, and including many functions that clearly should not be (and are not) system calls, such as an API call to copy an entire file. Many Win32 API calls create kernel objects of one kind or another, including files, processes, threads, pipes, etc. Every call creating a kernel object returns a result called a handle to the caller. This handle can be subsequently used to perform operations on the object. Handles are specific to the process that created the object referred to by the handle. They cannot be passed directly to another process and used there (just as UNIX file descriptors cannot be passed to other processes and used there). However, under certain circumstances, it is possible to duplicate a handle and pass it to other processes in a protected way, allowing them controlled access to objects belonging to other processes. Every object can have a security descriptor associated with it, telling in detail who may and may not perform what kinds of operations on the object. XP is sometimes said to be object-oriented because the only way to manipulate kernel objects is by invoking methods (API functions) on their handles. On the other hand, it lacks some of the most basic properties of object-oriented systems such as inheritance and polymorphism. The Win32 API was also available on Windows 95/98/ME (as well as on the consumer electronics operating system, Windows CE), with a small number of

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

479

exceptions. For example, Windows 95/98 did not have any security, so those API calls that relate to security just return error codes on Windows 95/98. Also, XP file names use the Unicode character set, which was not available on Windows 95/98. There are also differences in parameters to some API function calls. On XP, for example, all the screen coordinates given in the graphics functions are true 32-bit numbers; on Windows 95/98, only the low-order 16 bits were used (for backward compatibility with Windows 3.1). The existence of the Win32 API on multiple operating systems makes it easier to port programs between them but also points out more clearly that it is somewhat decoupled from the actual system calls.

6.4.2 Examples of Virtual Memory In this section we will look at virtual memory in both UNIX and XP. For the most part, they are fairly similar from the programmer’s point of view. UNIX Virtual Memory The UNIX memory model is simple. Each process has three segments: code, data, and stack, as illustrated in Fig. 6-32. In a machine with a single, linear address space, the code is generally placed near the bottom of memory, followed by the data. The stack is placed at the top of memory. The code size is fixed, but the data and stack may each grow, in opposite directions. This model is easy to implement on almost any machine and is the model used by Solaris. Address 0xFFFFFFFF

Stack

Data 0

Code

Figure 6-32. The address space of a single UNIX process.

Furthermore, if the machine has paging, the entire address space can be paged, without user programs even being aware of it. The only thing they notice is that it is permitted to have programs larger than the machine’s physical memory. UNIX systems that do not have paging generally swap entire processes between memory and disk to allow an arbitrarily large number of processes to be timeshared. For Berkeley UNIX , the above description (demand-paged virtual memory) is basically the entire story. However, System V (and also Solaris) include several

480

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

features that allow users to manage their virtual memory in more sophisticated ways. Most important of these is the ability of a process to map a (portion of a) file onto part of its address space. For example, if a 12-KB file is mapped at virtual address 144K, a read to the word at address 144 KB reads the first word of the file. In this way file I/O can be done without making system calls. Since some files may exceed the size of the virtual address space, it is also possible to map in only a portion of a file instead of the whole file. The mapping is done first opening the file and getting back a file descriptor, fd, which is used to identify the file to be mapped. Then the process makes a call paddr = mmap(virtual "address, length, protection, flags, fd, file"offset)

which maps length bytes starting at file"offset in the file onto the virtual address space starting at virtual"address. Alternatively, the flags parameter can be set to ask the system to choose a virtual address, which it then returns as paddr. The mapped region must be an integral number of pages and aligned at a page boundary. The protection parameter can specify any combination of read, write, or execute permission. The mapping can be removed later with unmap. Multiple processes can map onto the same file at the same time. Two options are provided for sharing. In the first one, all the pages are shared, so writes done by one process are visible to all the others. This option provides a high-bandwidth communication path between processes. The other option shares the pages as long as no process modifies them. However, as soon as any process attempts to write on a page, it gets a protection fault, which causes the operating system to give it a private copy of the page to write on. This scheme, known as copy on write, is used when each of multiple processes needs the illusion it is the only one mapped onto a file. Windows XP Virtual Memory In XP, every user process has its own virtual address space. Virtual addresses are 32 bits long, so each process has 4 GB of virtual address space. The lower 2 GB are available for the process’ code and data; the upper 2 GB allow (limited) access to kernel memory, except in Server versions of Windows, in which the split can be 3 GB for the user and 1 GB for the kernel. The virtual address space is demand paged, with a fixed page size (4 KB on the Pentium 4). Each virtual page can be in one of three states: free, reserved, or committed. A free page is not currently in use and a reference to it causes a page fault. When a process is started, all of its pages are in free state until the program and initial data are mapped into its address space. Once code or data is mapped onto a page, the page is said to be committed. A reference to a committed page is mapped using the virtual memory hardware and succeeds if the page is in main memory. If the page is not in main memory, a page fault occurs and the operating system finds and brings in the page from disk. A virtual page can also be in reserved

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

481

state, meaning it is not available for being mapped until the reservation is explicitly removed. In addition to the free, reserved, and committed attributes, pages also have other attributes, such as being readable, writable, and executable. The top 64 KB and bottom 64 KB of memory are always free, to catch pointer errors (uninitialized pointers are often 0 or −1). Each committed page has a shadow page on the disk where it is kept when it is not in main memory. Free and reserved pages do not have shadow pages, so references to them cause page faults (the system cannot bring in a page from disk if there is no page on disk). The shadow pages on the disk are arranged into one or more paging files. The operating system keeps track of which virtual page maps onto which part of which paging file. For (execute only) program text, the executable binary file contains the shadow pages; for data pages, special paging files are used. XP, like System V, allows files to be mapped directly onto regions of the virtual address spaces (i.e., runs of pages). Once a file has been mapped onto the address space, it can be read or written using ordinary memory references. Memory-mapped files are implemented in the same way as other committed pages, only the shadow pages can be in the disk file instead of in the paging file. As a result, when a file is mapped in, the version in memory may not be identical to the disk version (due to recent writes to the virtual address space). However, when the fileunmapped or is explicitly flushed, the disk version is updated. XP explicitly allows two or more processes to map in the same file at the same time, possibly at different virtual addresses. By reading and writing memory words, the processes can now communicate with each other and pass data back and forth at very high bandwidth, since no copying is required. Different processes may have different access permissions. Since all the processes using a mapped file share the same pages, changes made by one of them are immediately visible to all the others, even if the disk file has not yet been updated. The Win32 API contains a number of functions that allow a process to manage its virtual memory explicitly. The most important of these functions are listed in Fig. 6-33. All of them operate on a region consisting either of a single page or a sequence of two or more pages that are consecutive in the virtual address space. The first four API functions are self-explanatory. The next two give a process the ability to hardwire some numbe of pages in memory so they will not be paged out and to undo this property. A real-time program might need this ability, for example. Only programs run on behalf of the system administrator may pin pages in memory. And a limit is enforced by the operating system to prevent even these processes from getting too greedy. Although not shown in Fig. 6-33, XP also has API functions to allow a process to access the virtual memory of a different process over which it has been given control (i.e., for which it has a handle). The last four API functions listed are for managing memory-mapped files. To map a file, a file mapping object must first be created, with CreateFileMapping.

482

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

############################################################################### ! ! ! API function Meaning ############################################################################### ! ! ! ! VirtualAlloc ! Reserve or commit a region ! ############################################################################### ! VirtualFree ! Release or decommit a region ! ############################################################################### ! ! ! !############################################################################### ! Change the read/write/execute protection on a region ! VirtualProtect ! ! ! VirtualQuery Inquire about the status of a region !############################################################################### ! ! !############################################################################### ! Make a region memory resident (i.e., disable paging for it) ! VirtualLock ! ! ! VirtualUnlock !############################################################################### ! Make a region pageable in the usual way ! ! CreateFileMapping ! Create a file-mapping object and (optionally) assign it a name ! !############################################################################### ! ! MapViewOfFile !############################################################################### ! Map (part of) a file into the address space ! ! UnmapViewOfFile ! Remove a mapped file from the address space ! !############################################################################### ! ! !############################################################################### !! Open a previously created file mapping object !! ! OpenFileMapping Figure 6-33. The principal Windows XP API calls for managing virtual memory

This function returns a handle to the file mapping object and optionally enters a name for it into the file system so another process can use it. The next two functions map and unmap files, respectively. The last one can be used by a process to map in a file currently also mapped in by a different process. In this way, two or more processes can share regions of their address spaces. These API functions are the basic ones upon which the rest of the memory management system is built. For example, there are API functions for allocating and freeing data structures on one or more heaps. Heaps are used for storing data structures that are dynamically created and destroyed. The heaps are not garbage collected, so it is up to user software to free blocks of virtual memory that are no longer in use. (Garbage collection is the automatic removal of unused data structures by the system.) Heap usage in XP is similar to the use of the malloc function in UNIX systems, except that there can be multiple independently managed heaps.

6.4.3 Examples of Virtual I/O The heart of any operating system is providing services to user programs, mostly I/O services such as reading and writing files. Both UNIX and XP offer a wide variety of I/O services to user programs. For most UNIX system calls, XP has an equivalent call, but the reverse is not true, as XP has far more calls and each of them is far more complicated than its UNIX counterpart. UNIX Virtual I/O Much of the popularity of the UNIX system can be traced directly to its simplicity, which, in turn, is a direct result of the organization of the file system. An ordinary file is a linear sequence of 8-bit bytes starting at 0 and going up to a

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

483

maximum of 232 − 1 bytes. The operating system itself imposes no record structure on files, although many user programs regard ASCII text files as sequences of lines, each line terminated by a line feed. Associated with every open file is a pointer to the next byte to be read or written. The read and write system calls read and write data starting at the file position indicated by the pointer. Both calls advance the pointer after the operation by an amount equal to the number of bytes transferred. However, random access to files is possible by explicitly setting the file pointer to a specific value. In addition to ordinary files, the UNIX system also supports special files, which are used to access I/O devices. Each I/O device typically has one or more special files assigned to it. By reading and writing from the associated special file, a program can read or write from the I/O device. Disks, printers, terminals, and many other devices are handled this way. The major UNIX file system calls are listed in Fig. 6-34. The creat call (without the e) can be used to create a new file. It is not strictly necessary any more, because open can also create a new file now. Unlink removes a file, assuming that the file is in only one directory. ############################################################################## ! ! ! System call Meaning ############################################################################## ! ! ! ! creat(name, mode) ! Create a file; mode specifies the protection mode ! ############################################################################## ! unlink(name) ! Delete a file (assuming that there is only 1 link to it) ! ############################################################################## ! ! ! ! open(name, mode) ! Open or create a file and return a file descriptor ! ############################################################################## ! ! ! ############################################################################## ! close(fd) ! Close a file ! ! read(fd, buffer, count) ! Read count bytes into buffer ! ############################################################################## ! ! ! ! write(fd, buffer, count) ! Write count bytes from buffer ! ############################################################################## ! lseek(fd, offset, w) ! Move the file pointer as required by offset and w ! ############################################################################## ! ! ! ! stat(name, buffer) ! Return information about a file ! ############################################################################## ! chmod(name, mode) ! Change the protection mode of a file ! ############################################################################## ! ! ! !############################################################################## !! Do various control operations such as locking (part of) a file !! ! fcntl(fd, cmd, ...) Figure 6-34. The principal UNIX file system calls.

Open is used to open existing files (and create new ones). The mode flag tells how to open it (for reading, for writing, etc.). The call returns a small integer called a file descriptor that identifies the file in subsequent calls. When the file is no longer needed, close is called to free up the file descriptor. The actual file I/O is done with read and write, each of which has a file descriptor indicating which file to use, a buffer for the data to go to or come from, and a byte count telling how much data to transmit. Lseek is used to position the file pointer, making random access to files possible. Stat returns information about a file, including its size, time of last access, owner, and more. Chmod changes the protection mode of a file, for example,

484

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

allowing or forbidding users other than the owner from reading it. Finally, fcntl does various miscellaneous operations on a file, such as locking or unlocking it. Figure 6-35 illustrates how the major file I/O calls work. This code is minimal and does not include the necessary error checking. Before entering the loop, the program opens an existing file, data, and creates a new file, newf. Each call returns a file descriptor, infd, and outfd, respectively. The second parameters to the two calls are protection bits that specify that the files are to be read and written, respectively. Both calls return a file descriptor. If either open or creat fails, a negative file descriptor is returned, telling that the call failed. /* Open the file descriptors. */ infd = open(′′data′′, 0); outfd = creat(′′newf′′, ProtectionBits); /* Copy loop. */ do { count = read(infd, buffer, bytes); if (count > 0) write(outfd, buffer, count); } while (count > 0); /* Close the files. */ close(infd); close(outfd); Figure 6-35. A program fragment for copying a file using the UNIX system calls. This fragment is in C because Java hides the low-level system calls and we are trying to expose them.

The call to read has three parameters: a file descriptor, a buffer, and a byte count. The call tries to read the desired number of bytes from the indicated file into the buffer. The number of bytes actually read is returned in count, which will be smaller than bytes if the file was too short. The write call deposits the newly read bytes on the output file. The loop continues until the input file has been completely read, at which time the loop terminates and both files are closed. File descriptors in UNIX are small integers (usually below 20). File descriptors 0, 1, and 2 are special and correspond to standard input, standard output, and standard error, respectively. Normally, these refer to the keyboard, the display, and the display, respectively, but they can be redirected to files by the user. Many UNIX programs get their input from standard input and write the processed output on standard output. Such programs are often called filters. Closely related to the file system is the directory system. Each user may have multiple directories, with each directory containing both files and subdirectories. UNIX systems normally are configured with a main directory, called the root directory, containing subdirectories bin (for frequently executed programs), dev (for the special I/O device files), lib (for libraries), and usr (for user directories), as shown in Fig. 6-36. In this example, the usr directory contains subdirectories

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

485

for ast and jim. The ast directory contains two files, data and foo.c, and a subdirectory, bin, containing four games. Root directory bin dev lib usr

… /dev

/bin

…

/usr

…

/lib

ast jim

…

… /usr/ast

/usr/jim

bin

jotto

data foo.c

…

… /usr/ast/bin game 1 game 2 game 3 game 4

… Data files

Figure 6-36. Part of a typical UNIX directory system.

Files can be named by giving their path from the root directory. A path contains a list of all the directories traversed from the root to the file, with directory

486

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

names separated by slashes. For example, the absolute path name of game2 is /usr/ast/bin/game2. A path starting at the root is called an absolute path. At every instant, each running program has a working directory. Path names may also be relative to the working directory, in which case they do not begin with a slash, to distinguish them from absolute path names. Such paths are called relative paths. When /usr/ast is the working directory, game3 can be accessed using the path bin/game3. A user may create a link to someone else’s file using the link system call. In the above example, /usr/ast/bin/game3 and /usr/jim/jotto both access the same file. To prevent cycles in the directory system, links are not permitted to directories. The calls open and creat take either absolute or relative path names as arguments. The major directory management system calls in UNIX are listed in Fig. 6-37. Mkdir creates a new directory and rmdir deletes an existing (empty) directory. The next three calls are used to read directory entries. The first one opens the directory, the next one reads entries from it, and the last one closes the directory. Chdir changes the working directory. ######################################################################### ! ! System call Meaning # ! ######################################################################## ! ! mkdir(name, mode) ! Create a new directory ! #! ######################################################################## ! rmdir(name) ! Delete an empty directory ! #! ######################################################################## ! ! ! Open a directory for reading ! opendir(name) #! ######################################################################## ! ! ! readdir(dirpointer) !######################################################################### ! Read the next entry in a directory ! !######################################################################### ! Close a directory ! closedir(dirpointer) ! ! ! chdir(dirname) !######################################################################### ! Change working directory to dirname ! ! link(name1, name2) ! Create a directory entry name2 pointing to name1 ! !######################################################################### ! ! unlink(name) !######################################################################### ! Remove name from its directory ! !

Figure 6-37. The principal UNIX directory management calls.

Link makes a new directory entry with the new entry pointing to an existing file. For example, the entry /usr/jim/jotto might have been created by the call link(′′/usr/ast/bin/game3′′, ′′/usr/jim/jotto′′)

or an equivalent call using relative path names, depending on the working directory of the program making the call. Unlink removes a directory entry. If the file has only one link, the file is deleted. If it has two or more links, it is kept. It does not matter whether a removed link is the original or a copy made later. Once a link is made, it is a first-class citizen, indistinguishable from the original. The call unlink(′′/usr/ast/bin/game3′′)

makes game3 only accessible via the path /usr/jim/jotto henceforth. Link and unlink can be used in this way to ‘‘move’’ files from one directory to another.

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

487

Associated with every file (including directories, because they are also files) is a bit map telling who may access the file. The map contains three RWX fields, the first controlling the Read, Write, eXecute permissions for the owner, the second for others in the owner’s group, and the third for everybody else. Thus RWX R-X --X means that the owner can read the file, write the file, and execute the file (obviously, it is an executable program, or execute would be off), whereas others in his group can read or execute it and strangers can only execute it. With these permissions, strangers can use the program but not steal (copy) it because they do not have read permission. The assignment of users to groups is done by the system administrator, usually called the superuser. The superuser also has the power to override the protection mechanism and read, write, or execute any file. Let us now briefly examine how files and directories are implemented in UNIX. For a more complete treatment, see Vahalia (1996). Associated with each file (and each directory, because a directory is also a file) is a 64-byte block of information called an i-node. The i-node tells who owns the file, what the permissions are, where to find the data, and similar things. The i-nodes for the files on each disk are located either in numerical sequence at the beginning of the disk, or if the disk is split up into groups of cylinders, at the start of a cylinder group. Thus given an i-node number, the UNIX system can locate the i-node by simply calculating its disk address. A directory entry consists of two parts: a file name and an i-node number. When a program executes open(′′foo.c′′, 0)

the system searches the working directory for the file name, ‘‘foo.c,’’ in order to locate the i-node number for that file. Having found the i-node number, it can then read in the i-node, which tells it all about the file. When a longer path name is specified, the basic steps outlined above are repeated several times until the full path has been parsed. For example, to locate the i-node number for /usr/ast/data, the system first searches the root directory for an entry usr. Having found the i-node for usr, it can read that file (a directory is a file in UNIX). In this file it looks for an entry ast, thus locating the i-node number for the file /usr/ast. By reading /usr/ast, the system can then find the entry for data, and thus the i-node number for /usr/ast/data. Given the i-node number for the file, it can then find out everything about the file from the i-node. The format, contents, and layout of an i-node vary somewhat from system to system (especially when networking is in use), but the following items are typically found in each i-node. 1. The file type, the 9 RWX protection bits, and a few other bits. 2. The number of links to the file (number of directory entries for it). 3. The owner’s identity.

488

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

4. The owner’s group. 5. The file length in bytes. 6. Thirteen disk addresses. 7. The time the file was last read. 8. The time the file was last written. 9. The time the i-node was last changed. The file type distinguishes ordinary files, directories, and two kinds of special files, for block-structured and unstructured I/O devices, respectively. The number of links and the owner identification have already been discussed. The file length is a 32-bit integer giving the highest byte that has a value. It is perfectly legal to create a file, do an lseek to position 1,000,000, and write 1 byte, which yields a file of length 1,000,001. The file would not, however, require storage for all the ‘‘missing’’ bytes. The first 10 disk addresses point to data blocks. With a block size of 1024 bytes, files up to 10,240 bytes can be handled this way. Address 11 points to a disk block, called an indirect block, which contains 256 disk addresses. Files up to 10,240 + 256 × 1024 = 272,384 bytes are handled this way. For still larger files, address 12 points to a block containing the addresses of 256 indirect blocks, which takes care of files up to 272,384 + 256 × 256 × 1024 = 67,381,248 bytes. If this double indirect block scheme is still too small, disk address 13 is used to point to a triple indirect block containing the addresses of 256 double indirect blocks. Using the direct, single, double, and triple indirect addresses, up to 16,843,018 blocks can be addressed giving a theoretical maximum file size of 17,247,250,432 bytes. Since file pointers are limited to 32 bits, the practical upper limit is actually 4,294,967,295 bytes. Free disk blocks are kept on a linked list. When a new block is needed, the next block is plucked from the list. As a result, the blocks of each file are scattered randomly around the disk. To make disk I/O more efficient, when a file is opened, its i-node is copied to a table in main memory and is kept there for handy reference as long as the file remains open. In addition, a pool of recently referenced disk blocks is maintained in memory. Because most files are read sequentially, it often happens that a file reference requires the same disk block as the previous reference. To strengthen this effect, the system also tries to read the next block in a file, before it is referenced, in order to speed up processing. All this optimization is hidden from the user; when a user issues a read call, the program is suspended until the requested data are available in the buffer. With this background information, we can now take a look to see how file I/O works. Open causes the system to search the directories for the specified path. If the search is successful, the i-node is read into an internal table. Reads and writes require the system to compute the block number from the current file position.

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

489

The disk addresses of the first 10 blocks are always in main memory (in the inode); higher-numbered blocks require one or more indirect blocks to be read first. Lseek just changes the current position pointer without doing any I/O. Link and unlink are also simple to understand now. Link looks up its first argument to find the i-node number. Then it creates a directory entry for the second argument, putting the i-node number of the first file in that entry. Finally, it increases the link count in the i-node by one. Unlink removes a directory entry and decrements the link count in the i-node. If it is zero, the file is removed and all the blocks are put back on the free list. Windows XP Virtual I/O XP supports several file systems, the most important of which are NTFS (NT File System) and the FAT (File Allocation Table) file system. The former is a new file system developed specifically for XP; the latter is the old MS-DOS file system, which was also used on Windows 95/98 (albeit with support for longer file names). Since the FAT file system is basically obsolete, we will study NTFS below. File names in NTFS can be up to 255 characters long. File names are in Unicode, allowing people in countries not using the Latin alphabet (e.g., Japan, India, and Israel) to write file names in their native language. (In fact, XP uses Unicode throughout internally; versions starting with Windows 2000 have a single binary that can be used in any country and still use the local language because all the menus, error messages, etc., are kept in country-dependent configuration files.) NTFS fully supports case-sensitive names (so foo is different from FOO). Unfortunately, the Win32 API does not fully support case-sensitivity for file names and not at all for directory names, so this advantage is lost to programs using Win32. As with UNIX, a file is just a linear sequence of bytes, although up to a maximum of 264 − 1. File pointers also exist, as in UNIX , but are 64 bits wide rather than 32 bits, to handle the maximum length file. The Win32 API function calls for file and directory manipulation are roughly similar to their UNIX counterparts, except most have more parameters and the security model is different. Opening a file returns a handle, which is then used for reading and writing the file. However, unlike in UNIX, handles are not small integers, and standard input, standard output, and standard error have to be acquired explicitly rather than being predefined as 0, 1, and 2 (except in console mode, where they are preopened). The principal Win32 API functions for file management are listed in Fig. 6-38. Let us now examine these calls briefly. CreateFile can be used to create a new file and return a handle to it. This API function is also used to open existing files as there is no open API function. We have not listed the parameters for the XP API functions because they are so voluminous. As an example, CreateFile has seven parameters, as follows:

490

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

############################################################################### ! ! UNIX ! ! API function Meaning ############################################################################### ! ! ! ! ! CreateFile ! open ! Create a file or open an existing file; return a handle ! ############################################################################### ! DeleteFile ! unlink ! Destroy an existing file ! ############################################################################### ! ! ! ! ! CloseHandle ! close ! Close a file ! ############################################################################### ! ! ! ! ReadFile read Read data from a file ############################################################################### ! ! ! ! ! WriteFile ! write ! Write data to a file ! ############################################################################### ! ! ! ! ! SetFilePointer ! lseek ! Set the file pointer to a specific place in the file ! ############################################################################### ! GetFileAttributes ! stat ! Return the file properties ! ############################################################################### ! ! ! ! ! LockFile ! fcntl ! Lock a region of the file to provide mutual exclusion ! ############################################################################### ! UnlockFile ! fcntl ! Unlock a previously locked region of the file ! ! ############################################################################### ! ! ! Figure 6-38. The principal Win32 API functions for file I/O. The second column gives the nearest UNIX equivalent.

1. A pointer to the name of the file to create or open. 2. Flags telling whether the file can be read, written, or both. 3. Flags telling whether multiple processes can open the file at once. 4. A pointer to the security descriptor, telling who can access the file. 5. Flags telling what to do if the file exists/does not exist. 6. Flags dealing with attributes such as archiving, compression, etc. 7. The handle of a file whose attributes should be cloned for the new file. The next six API functions in Fig. 6-38 are fairly similar to the corresponding

UNIX system calls. The last two allow a region of a file to be locked and

unlocked to permit a process to get guaranteed mutual exclusion to it. Using these API functions, it is possible to write a procedure to copy a file, analogous to the UNIX version of Figure 6-35. Such a procedure (without any error checking) is shown in Fig. 6-39. It has been designed to mimic the structure of Figure 6-35. In practice, one would not have to program a copy file program since CopyFile is an API function (which executes something close to this program as a library procedure). XP supports a hierarchical file system, similar to the UNIX file system. The separator between component names is \ however, instead of /, a fossil inherited from MS-DOS. There is a concept of a current working directory and path names can be relative or absolute. One significant difference with UNIX , however, is that UNIX allows the file systems on different disks and machines to be mounted together in a single naming tree, thus hiding the disk structure from all software. XP 4.0 does not have this property, so absolute file names must begin with a drive

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

491

/* Open files for input and output. */ inhandle = CreateFile(′′data′′, GENERIC "READ, 0, NULL, OPEN"EXISTING, 0, NULL); outhandle = CreateFile(′′newf′′, GENERIC "WRITE, 0, NULL, CREATE "ALWAYS, FILE"ATTRIBUTE "NORMAL, NULL); /* Copy the file. */ do { s = ReadFile(inhandle, buffer, BUF"SIZE, &count, NULL); if (s > 0 && count > 0) WriteFile(outhandle, buffer, count, &ocnt, NULL); } while (s > 0 && count > 0); /* Close the files. */ CloseHandle(inhandle); CloseHandle(outhandle); Figure 6-39. A program fragment for copying a file using the Windows XP API functions. This fragment is in C because Java hides the low-level system calls and we are trying to expose them.

letter indicating which logical disk is meant, as in C:\windows\system\foo.dll. Starting with Windows 2000 UNIX-style mounting of file systems was added. The major directory management API functions are given in Fig. 6-40, again along with their nearest UNIX equivalents. The functions are hopefully selfexplanatory. ############################################################################### ! UNIX ! ! API function Meaning ! ! ! !############################################################################### CreateDirectory ! mkdir ! Create a new directory ! !############################################################################### ! rmdir ! Remove an empty directory ! ! RemoveDirectory ! ! ! !############################################################################### ! opendir ! Initialize to start reading the entries in a directory ! !############################################################################### FindFirstFile ! ! ! ! FindNextFile ! readdir ! Read the next directory entry ! !############################################################################### ! ! Move a file from one directory to another ! !############################################################################### MoveFile ! ! ! ! SetCurrentDirectory ! chdir ! Change the current working directory ! !############################################################################### !

Figure 6-40. The principal Win32 API functions for directory management. The second column gives the nearest UNIX equivalent, when one exists.

XP has a much more elaborate security mechanism than most UNIX systems. Although there are hundreds of API functions relating to security, the following brief description gives the general idea. When a user logs in, his or her initial process is given an access token by the operating system. The access token contains the user’s SID (Security ID), a list of the security groups to which the user belongs, any special privileges available, and a few other items. The point of the access token is to concentrate all the security information in one easy-to-find place. All processes created by this process inherit the same access token.

492

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

One of the parameters that can be supplied when any object is created is its security descriptor. The security descriptor contains a list of entries called an ACL (Access Control List). Each entry permits or prohibits some set of the operations on the object by some SID or group. For example, a file could have a security descriptor specifying that Elinor has no access to the file at all, Ken can read the file, Linda can read or write the file, and that all members of the XYZ group can read the file’s length but nothing else. When a process tries to perform some operation on an object using the handle it got when it opened the object, the security manager gets the process’ access token and goes down the list of entries in the ACL in order. As soon as it finds an entry that matches the caller’s SID or one of the caller’s groups, the access found there is taken as definitive. For this reason, it is usual to put entries denying access ahead of entries granting access in the ACL, so that a user who is specifically denied access cannot get in via a back door by being a member of a group that has legitimate access. The security descriptor also contains information used for auditing accesses to the object. Let us now take a quick look at how files and directories are implemented in XP. Each disk is statically divided up into self-contained volumes, which are the same as disk partitions in UNIX. Each volume contains bit maps, files, directories, and other data structures for managing its information. Each volume is organized as a linear sequence of clusters, with the cluster size being fixed for each volume and ranging from 512 bytes to 64 KB, depending on the volume size. Clusters are referred to by their offset from the start of the volume using 64-bit numbers. The main data structure in each volume is the MFT (Master File Table), which has an entry for each file and directory in the volume. These entries are analogous to the i-nodes in UNIX. The MFT is itself a file, and as such can be placed anywhere within the volume, thus eliminating the problem that UNIX has with bad disk blocks in the middle of the i-nodes. The MFT is shown in Fig. 6-41. It begins with a header containing information about the volume, such as (pointers to) the root directory, the boot file, the bad-block file, the free-list administration, etc. After that comes an entry per file or directory, 1 KB except when the cluster size is 2 KB or more. Each entry contains all the metadata (administrative information) about the file or directory. Several formats are allowed, one of which is shown in Fig. 6-41. The standard information field contains information such as the time stamps needed by the hard link count, the read-only and archive bits, etc. It is a fixedlength field and always present. The file name is variable length, up to 255 Unicode characters. In order to make such files accessible to old 16-bit programs, files can also have a MS-DOS name, which consists of eight alphanumeric characters optionally followed by a dot and an extension of up to three alphanumeric characters. If the actual file name conforms to the MS-DOS 8+3 naming rule, a secondary MS-DOS name is not used.

SEC. 6.4

493

EXAMPLE OPERATING SYSTEMS

Standard MS-DOS information File name name Security

Data

MFT entry for one file MFT header Master file table

Figure 6-41. The Windows XP master file table.

Next comes the security information. In versions up to and including XP 4.0, the security field contained the actual security descriptor. Starting with Windows 2000, all the security information was centralized in a single file, with the security field simply pointing to the relevant part of this file. For small files, the file data itself is actually contained in the MFT entry, saving a disk access to fetch it. This idea is called an immediate file (Mullender and Tanenbaum, 1984). For somewhat larger files, this field contains pointers to the clusters containing the data, or more commonly, runs of consecutive clusters so a single cluster number and a length can represent an arbitrary amount of file data. If a single MFT entry is insufficiently large to hold whatever information it is supposed to hold, one or more additional entries can be chained to it. The maximum file size is 264 bytes. To get an idea of how big a 264 -byte file is, imagine that the file were written out in binary, with each 0 or 1 occupying 1 mm of space. The 267 -mm listing would be 15 light-years long, reaching far beyond the solar system, to Alpha Centauri and back. The NTFS file system has many other interesting properties including data compression and fault tolerance using atomic transactions. Additional information about it can be found in (Russinovich and Solomon 2005).

6.4.4 Examples of Process Management Both UNIX and XP allow a job to be split up into multiple processes that can run in (pseudo)parallel and communicate with each other, in the style of the producer-consumer example discussed earlier. In this section we will discuss how processes are managed in both systems. Both systems also support parallelism within a single process using threads, so that will also be discussed.

494

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

UNIX Process Management At any time, a UNIX process can create a subprocess that is an exact replica of itself by executing the fork system call. The original process is called the parent and the new one is called the child. Right after the fork, the two processes are identical and even share the same file descriptors. Thereafter, each one goes its own way and does whatever it wants to, independent of the other one. In many cases, the child process juggles the file descriptors in certain ways and then executes the exec system call, which replaces its program and data with the program and data found in an executable file specified as parameter to the exec call. For example, when a user types a command xyz at a terminal, the command interpreter (shell) executes fork to create a child process. This child process then executes exec to run the xyz program. The two processes run in parallel (with or without exec), unless the parent wishes to wait for the child to terminate before continuing. If the parent wishes to wait, it executes either the wait or waitpid system call, which causes it to be suspended until the child finishes by executing exit. After the child finishes, the parent continues. Processes can execute fork as often as they want, giving rise to a tree of processes. In Fig. 6-42, for example, process A has executed fork twice, creating two children, B and C. Then B also executed fork twice, and C executed it once, giving the final tree of six processes. A

B

D

E

Original process

C

Children of A

F

Grandchildren of A

Figure 6-42. A process tree in UNIX.

Processes in UNIX can communicate with each other via a structure called a pipe. A pipe is a kind of buffer into which one process can write a stream of data and another can take it out. Bytes are always retrieved from a pipe in the order they were written. Random access is not possible. Pipes do not preserve message boundaries, so if one process does four 128-byte writes and the other does a 512byte read, the reader will get all the data at once, with no indication that they were written in multiple operations. In System V and Solaris, another way for processes to communicate is by using message queues. A process can create a new message queue or open an existing one using msgget. Using a message queue, a process can send messages

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

495

using msgsnd and receive them using msgrecv. Messages sent this way differ from data stuffed into a pipe in several ways. First, message boundaries are preserved, whereas a pipe is just a byte stream. Second, messages have priorities, so urgent ones can skip ahead of less important ones. Third, messages are typed, and a msgrecv can specify a particular type, if desired. Another communication mechanism is the ability of two or more processes to share a region of their respective address spaces. UNIX handles this shared memory by mapping the same pages into the virtual address space of all the sharing processes. As a result, a write by one process into the shared region is immediately visible to the other processes. This mechanism provides a very high bandwidth communication path between processes. The system calls involved in shared memory go by names like shmat and shmop. Another feature of System V and Solaris is the availability of semaphores. These work essentially as described in the producer-consumer example given in the text. Another facility provided by all UNIX systems is the ability to have multiple threads of control within a single process. These threads of control, usually just called threads, are like lightweight processes that share a common address space and everything associated with that address space, such as file descriptors, environment variables, and outstanding timers. However, each thread has its own program counter, own registers, and own stack. When a thread blocks (i.e., has to stop temporarily until I/O completes or some other event happens), other threads in the same process are still able to run. Two threads in the same process operating as a producer and consumer are similar, but not identical, to two single-thread processes that are sharing a memory segment containing a buffer. The differences have to do with the fact that in the latter case, each process has its own file descriptors, etc., whereas in the former case all of these items are shared. We saw the use of Java threads in our producer-consumer example earlier. Often the Java runtime system uses an operating system thread for each of its threads, but it does not have to do this. As an example of where threads might be useful, consider a World Wide Web server. Such a server might keep a cache of commonly-used Web pages in main memory. If a request is for a page in the cache, the Web page is returned immediately. Otherwise, it is fetched from disk. Unfortunately, waiting for the disk takes a long time (typically 20 msec), during which the process is blocked and cannot serve new incoming requests, even those for Web pages in the cache. The solution is to have multiple threads within the server process, all of which share the common Web page cache. When one thread blocks, other threads can handle new requests. To prevent blocking without threads, one could have multiple server processes, but this would probably entail replicating the cache, thus wasting valuable memory. The UNIX standard for threads is called pthreads, and is defined by (P1003.1C). It contains calls for managing and synchronizing threads. It is not

496

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

defined whether threads are managed by the kernel or entirely in user space. The most commonly-used thread calls are listed in Fig. 6-43. ############################################################################### ! ! ! Thread call Meaning ############################################################################### ! ! ! ! pthread"create ! Create a new thread in the caller’s address space ! ############################################################################### ! pthread"exit ! Terminate the calling thread ! ############################################################################### ! ! ! ! pthread"join ! Wait for a thread to terminate ! ############################################################################### ! ! ! ############################################################################### ! pthread"mutex"init ! Create a new mutex ! !############################################################################### ! Destroy a mutex ! pthread"mutex"destroy ! ! ! ! pthread"mutex"lock ! Lock a mutex ! ############################################################################### ! pthread"mutex"unlock ! Unlock a mutex ! ############################################################################### ! ! ! ! pthread"cond"init ! Create a condition variable ! ############################################################################### ! pthread"cond"destroy ! Destroy a condition variable ! ############################################################################### ! ! ! ! pthread"cond"wait ! Wait on a condition variable ! ############################################################################### ! ! ! pthread " cond " signal Release one thread waiting on a condition variable ! ############################################################################### ! ! Figure 6-43. The principal POSIX thread calls.

Let us briefly examine the thread calls shown in Fig. 6-43. The first call,

pthread"create, creates a new thread. After successful completion, one more

thread is running in the caller’s address space than was before the call. A thread that has done its job and wants to terminate calls pthread"exit. A thread can wait for another thread to exit by calling pthread"join. If the thread waited for has already exited, the pthread"join finishes immediately. Otherwise it blocks. Threads can synchronize using locks called mutexes. Typically a mutex guards some resource, such as a buffer shared by two threads. To make sure that only one thread at a time accesses the shared resource, threads are expected to lock the mutex before touching the resource and unlock it when they are done. As long as all threads obey this protocol, race conditions can be avoided. Mutexes are like binary semaphores, that is, semaphores that can take on only the values of 0 and 1. The name ‘‘mutex’’ comes from the fact that mutexes are used to ensure mutual exclusion on some resource. Mutexes can be created and destroyed by the calls pthread"mutex"init and pthread"mutex"destroy, respectively. A mutex can be in one of two states: locked or unlocked. When a thread needs to set a lock on an unlocked mutex (using pthread"mutex"lock), the lock is set and the thread continues. However, when a thread tries to lock a mutex that is already locked, it blocks. When the locking thread is finished with the shared resource, it is expected to unlock the corresponding mutex by calling pthread"mutex"unlock. Mutexes are intended for short-term locking, such as protecting a shared variable. They are not intended for long-term synchronization, such as waiting for a tape drive to become free. For long-term synchronization, condition variables

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

497

are provided. These are created and destroyed by calls to pthread"cond"init and pthread"cond"destroy, respectively. A condition variable is used by having one thread wait on it, and another signal it. For example, having discovered that the tape drive it needs is busy, a thread would do pthread"cond"wait on a condition variable that all the threads have agreed to associate with the tape drive. When the thread using the tape drive is finally done with it (possibly hours later), it uses pthread"cond"signal to release exactly one thread waiting on that condition variable (if any). If no thread is waiting, the signal is lost. Condition variables do not count like semaphores. A few other operations are also defined on threads, mutexes, and condition variables. Windows XP Process Management XP supports multiple processes, which can communicate and synchronize. Each process contains at least one thread, which in turn contains at least one fiber (lightweight thread). Together, processes, threads, and fibers provide a very general set of tools for managing parallelism, both on uniprocessors (single-CPU machines) and on multiprocessors (multi-CPU machines). New processes are created using the API function CreateProcess . This function has 10 parameters, each of which has many options. This design is clearly a lot more complicated that the UNIX scheme, in which fork has no parameters, and exec has just three: pointers to the name of the file to execute, the (parsed) command line parameter array, and the environment strings. Roughly speaking, the 10 parameters to CreateProcess are as follows:

1. A pointer to the name of the executable file. 2. The command line itself (unparsed). 3. A pointer to a security descriptor for the process. 4. A pointer to a security descriptor for the initial thread. 5. A bit telling whether the new process inherits the creator’s handles. 6. Miscellaneous flags (e.g., error mode, priority, debugging, consoles). 7. A pointer to the environment strings. 8. A pointer to the name of the new process’ current working directory. 9. A pointer to a structure describing the initial window on the screen. 10. A pointer to a structure that returns 18 values to the caller. XP does not enforce any kind of parent-child or other hierarchy. All processes are created equal. However, since 1 of the 18 parameters returned to the creating process is a handle to the new process (allowing considerable control over the

498

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

new process), there is an implicit hierarchy in terms of who has a handle to whom. Although these handles cannot just be passed directly to other processes, there is a way for a process to make a handle suitable for another process and then give it the handle, so the implicit process hierarchy may not last long. Each process in XP is created with a single thread, but a process can create more threads later on. Thread creation is simpler than process creation: CreateThread has only six parameters instead of 10: the security descriptor, the stack size, the starting address, a user-defined parameter, the initial state of the thread (ready or blocked), and the thread’s ID. The kernel does the thread creation, so it is clearly aware of threads (i.e., they are not implemented purely in user space as is the case in some other systems). When the kernel does scheduling, it not only picks the process to run next but also which thread in that process. This means that the kernel is always aware of which threads are ready and which ones are blocked. Because threads are kernel objects, they have security descriptors and handles. Since a handle for a thread can be passed to another process, it is possible to have one process control the threads in a different process. This feature is useful for debuggers, for example. Threads in XP are relatively expensive because doing a thread switch requires entering and later leaving the kernel. To provide very lightweight pseudoparallelism, XP provides fibers, which are like threads, but are scheduled in user space by the program that created them (or its runtime system). Each thread can have multiple fibers, the same way a process can have multiple threads, except that when a fiber logically blocks, it puts itself on the queue of blocked fibers and selects another fiber to run in the context of its thread. The kernel is not aware of this transition because the thread keeps running, even though it may be first running one fiber, then another. The kernel just manages processes and threads, not fibers. Fibers are useful, for example, when programs that manage their own threads are ported to XP. Processes can communicate in a wide variety of ways, including pipes, named pipes, mailslots, sockets, remote procedure calls, and shared files. Pipes have two modes: byte and message, selected at creation time. Byte-mode pipes work the same way as in UNIX. Message-mode pipes are somewhat similar but preserve message boundaries, so that four writes of 128 bytes will be read as four 128-byte messages, and not as one 512-byte message, as would happen with byte-mode pipes. Named pipes also exist and have the same two modes as regular pipes. Named pipes can also be used over a network; regular pipes cannot. Mailslots are a feature of XP not present in UNIX. They are similar to pipes in some ways, but not all. For one thing, they are one-way, whereas pipes are two-way. They can also be used over a network but do not provide guaranteed delivery. Finally, they allow the sending process to broadcast a message to many receivers, instead of to just one. Sockets are like pipes, except that they normally connect processes on different machines. However, they can also be used to connect processes on the

SEC. 6.4

EXAMPLE OPERATING SYSTEMS

499

same machine. In general, there is usually little advantage to using a socket connection over a pipe or named pipe for intramachine communication. Remote procedure calls are a way for process A to have process B call a procedure in B’s address space on A’s behalf and return the result to A. Various restrictions on the parameters exist. For example, it makes no sense to pass a pointer to a different process. Finally, processes can share memory by mapping onto the same file at the same time. All writes done by one process then appear in the address spaces of the other processes. Using this mechanism, the shared buffer used in our producer-consumer example can be easily implemented. Just as XP provides numerous interprocess communication mechanisms, it also provides numerous synchronization mechanisms, including semaphores, mutexes, critical sections, and events. All of these mechanisms work on threads, not processes, so that when a thread blocks on a semaphore, other threads in that process (if any) are not affected and can continue to run. A semaphore is created using the CreateSemaphore API function, which can initialize it to a given value and define a maximum value as well. Semaphores are kernel objects and thus have security descriptors and handles. The handle for a semaphore can be duplicated using DuplicateHandle and passed to another process so that multiple processes can synchronize on the same semaphore. Calls for up and down are present, although they have the somewhat peculiar names of ReleaseSemaphore (up) and WaitForSingleObject (down). It is also possible to give WaitForSingleObject a timeout, so the calling thread can be released eventually, even if the semaphore remains at 0 (although timers reintroduce races). Mutexes are also kernel objects used for synchronization, but simpler than semaphores because they do not have counters. They are essentially locks, with API functions for locking (WaitForSingleObject ) and unlocking (ReleaseMutex ). Like semaphore handles, mutex handles can be duplicated and passed between processes so that threads in different processes can access the same mutex. The third synchronization mechanism is based on critical sections, which are similar to mutexes, except local to the address space of the creating thread. Because critical sections are not kernel objects, they do not have handles or security descriptors and cannot be passed between processes. Locking and unlocking is done with EnterCriticalSection and LeaveCriticalSection, respectively. Because these API functions are performed entirely in user space, they are much faster than mutexes. The last synchronization mechanism uses kernel objects called events. A thread can wait for an event to occur with WaitForSingleObject . A thread can release a single thread waiting on an event with SetEvent or it can release all threads waiting on an event with PulseEvent. Events come in several flavors and have a variety of options, too. Events, mutexes, and semaphores can all be named and stored in the file system, like named pipes. Two or more processes can synchronize by opening the

500

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

same event, mutex, or semaphore, rather than having one of them create the object and then make duplicate handles for the others, although the latter approach is certainly an option as well.

6.5 SUMMARY The operating system can be regarded as an interpreter for certain architectural features not found at the ISA level. Chief among these are virtual memory, virtual I/O instructions, and facilities for parallel processing. Virtual memory is an architectural feature whose purpose is to allow programs to use more address space than the machine has physical memory, or to provide a consistent and flexible mechanism for memory protection and sharing. It can be implemented as pure paging, pure segmentation, or a combination of the two. In pure paging, the address space is broken up into equal-sized virtual pages. Some of these are mapped onto physical page frames. Others are not mapped. A reference to a mapped page is translated by the MMU into the correct physical address. A reference to an unmapped page causes a page fault. Both the Pentium 4 and the UltraSPARC III have MMUs that support virtual memory and paging. The most important I/O abstraction present at this level is the file. A file consists of a sequence of bytes or logical records that can be read and written without knowledge of how disks, tapes, and other I/O devices work. Files can be accessed sequentially, randomly by record number, or randomly by key. Directories can be used to group files together. Files can be stored in consecutive sectors or scattered around the disk. In the latter case, normal on hard disks, data structures are needed to locate all the blocks of a file. Free disk storage can be kept track of using a list or a bit map. Parallel processing is often supported and is implemented by simulating multiple processors by timesharing a single CPU. Uncontrolled interaction between processes can lead to race conditions. To solve this problem, synchronization primitives are introduced, of which semaphores are a simple example. Using semaphores, producer-consumer problems can be solved simply and elegantly. Two examples of sophisticated operating systems are UNIX and XP. Both support paging and memory-mapped files. They also both support hierarchical file systems, with files consisting of byte sequences. Finally, both support processes and threads and provide ways to synchronize them. PROBLEMS 1. Why does an operating system interpret only some of the level 3 instructions, whereas a microprogram interprets all the ISA level instructions? 2. A machine has a 32-bit byte-addressable virtual address space. The page size is 4 KB. How many pages of virtual address space exist?

CHAP. 6

PROBLEMS

501

3. Is it necessary to have the page size be a power of 2? Could a page of size, say, 4000 bytes be implemented in theory? If so, would it be practical? 4. A virtual memory has a page size of 1024 words, eight virtual pages, and four physical page frames. The page table is as follows: ################################ !################################ ! Virtual page ! Page frame ! ! ! 0 3 !################################ ! ! 1 1 !################################ ! ! ! ! not in main memory ! 2 ################################ ! ! ! 3 !################################ ! not in main memory ! !################################ ! ! 4 2 ! ! not in main memory ! 5 !################################ ! ! 6 0 !################################ ! ! 7 !! ################################ !! not in main memory !! a. Make a list of all virtual addresses that will cause page faults. b. What are the physical addresses for 0, 3728, 1023, 1024, 1025, 7800, and 4096? 5. A computer has 16 pages of virtual address space but only four page frames. Initially, the memory is empty. A program references the virtual pages in the order 0, 7, 2, 7, 5, 8, 9, 2, 4 a. Which references cause a page fault with LRU? b. Which references cause a page fault with FIFO? 6. In Sec. 6.1.4 an algorithm was presented for implementing a FIFO page replacement strategy. Devise a more efficient one. Hint: It is possible to update the counter in the newly-loaded page, leaving all the others alone. 7. In the paged systems discussed in the text, the page fault handler was part of the ISA level and thus was not present in any OSM level program’s address space. In reality, the page fault handler also occupies pages, and might, under some circumstances (e.g., FIFO page replacement policy), itself be removed. What would happen if the page fault handler were not present when a page fault occurred? How could this be fixed? 8. Not all computers have a hardware bit that is automatically set when a page is written to. Nevertheless, it is useful to keep track of which pages have been modified, to avoid having to assume worst case and write all pages back to the disk after use. Assuming that each page has hardware bits to separately enable access for reading, writing, and execution, how can the operating system keep track of which pages are clean and which are dirty? 9. A segmented memory has paged segments. Each virtual address has a 2-bit segment number, a 2-bit page number, and an 11-bit offset within the page. The main memory contains 32 KB, divided up into 2-KB pages. Each segment is either read-only, read/execute, read/write, or read/write/execute. The page tables and protection are as follows:

502

CHAP. 6 THE OPERATING SYSTEM MACHINE LEVEL ######################################################################## !######################################################################## ! ! ! ! Segment 0 Segment 1 Segment 2 Segment 3 ! ! ! ! ! Read only Read/execute Read/write/execute Read/write !######################################################################## ! ! ! ! Page Page Page ! ! Virtual ! Virtual ! ! Virtual page frame ! page frame ! frame ! !######################################################################## ! page ! ! ! ! 0 9 0 On disk ! Page table 0 14 ! ! ! ! ! 1 3 1 0 not in 1 1 ! ! ! ! ! 2 On disk 2 15 main 2 6 ! ! ! ! ! 3 12 3 8 memory 3 On disk ! !######################################################################## ! ! ! For each of the following accesses to virtual memory, tell what physical address is computed. If a fault occurs, tell which kind. Access

1. fetch data 2. fetch data 3. fetch data 4. store data 5. store data 6. store data 7. branch to it 8. fetch data 9. fetch data 10. branch to it

Segment 0 1 3 0 3 3 1 0 2 3

Page 1 1 3 1 1 0 3 2 0 0

Offset within page 1 10 2047 4 2 14 100 50 5 60

10. Some computers allow I/O directly to user space. For example, a program could start up a disk transfer to a buffer inside a user process. Does this cause any problems if compaction is used to implement the virtual memory? Discuss. 11. Operating systems that allow memory-mapped files always require files to be mapped at page boundaries. For example, with 4-KB pages, a file can be mapped in starting at virtual address 4096, but not starting at virtual address 5000. Why? 12. When a segment register is loaded on the Pentium 4, the corresponding descriptor is fetched and loaded into an invisible part of the segment register. Why do you think the Intel designers decided to do this? 13. A program on the Pentium 4 references local segment 10 with offset 8000. The BASE field of LDT segment 10 contains 10000. Which page directory entry does the Pentium 4 use? What is the page number? What is the offset? 14. Discuss some possible algorithms for removing segments in an unpaged, segmented memory. 15. Compare internal fragmentation to external fragmentation. What can be done to alleviate each? 16. Supermarkets are constantly faced with a problem similar to page replacement in virtual memory systems. They have a fixed amount of shelf space to display an everincreasing number of products. If an important new product comes along, say, 100%

CHAP. 6

PROBLEMS

503

efficient dog food, some existing product must be dropped from the inventory to make room for it. The obvious replacement algorithms are LRU and FIFO. Which of these would you prefer? 17. In some ways, caching and paging are very similar. In both cases there are two levels of memory (the cache and main memory in the former and main memory and disk in the latter). In this chapter we looked at some of the arguments in favor of large disk pages and small disk pages. Do the same arguments hold for cache line sizes? 18. Why do many file systems require that a file be explicitly opened with an open system call before being read? 19. Compare the bit-map and hole-list methods for keeping track of free space on a disk with 800 cylinders, each one having 5 tracks of 32 sectors. How many holes would it take before the hole list would be larger than the bit map? Assume that the allocation unit is the sector and that a hole requires a 32-bit table entry. 20. To be able to make some predictions of disk performance, it is useful to have a model of storage allocation. Suppose that the disk is viewed as a linear address space of N >> 1 sectors, consisting of a run of data blocks, then a hole, then another run of data blocks, and so on. If empirical measurements show that the probability distributions for data and hole lengths are the same, with the chance of either being i sectors as 2−i , what is the expected number of holes on the disk? 21. On a certain computer, a program can create as many files as it needs, and all files may grow dynamically during execution without giving the operating system any advance information about their ultimate size. Do you think that files are stored in consecutive sectors? Explain. 22. Studies of different file systems have shown that more than half the files are a few KB or smaller, with the vast majority of files less than something like 8 KB. On the other hand, the largest 10 percent of all files usually occupies about 95 percent of the entire disk space in use. From this data, what conclusion can you draw about disk block size? 23. Consider the following method by which an operating system might implement semaphore instructions. Whenever the CPU is about to do an up or down on a semaphore (an integer variable in memory), it first sets the CPU priority or mask bits in such a way as to disable all interrupts. Then it fetches the semaphore, modifies it, and branches accordingly. Finally, it enables interrupts again. Does this method work if a. There is a single CPU that switches between processes every 100 msec? b. Two CPUs share a common memory in which the semaphore is located? 24. The Nevercrash Operating System Company has been receiving complaints from some of its customers about its latest release, which includes semaphore operations. They feel it is immoral for processes to block (they call it ‘‘sleeping on the job’’). Since it is company policy to give the customers what they want, it has been proposed to add a third operation, peek, to supplement up and down. peek simply examines the semaphore without changing it or blocking the process. In this way, programs that feel it is immoral to block can first inspect the semaphore to see if it is safe to do a down. Will this idea work if three or more processes use the semaphore? If two processes use the semaphore?

504

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

25. Make a table showing which of the processes P1, P2, and P3 are running and which are blocked as a function of time from 0 to 1000 msec. All three processes perform up and down instructions on the same semaphore. When two processes are blocked and an up is done, the process with the lower number is restarted, that is, P1 gets preference over P2 and P3, and so on. Initially, all three are running and the semaphore is 1. At t = 100 At t = 200 At t = 300 At t = 400 At t = 500 At t = 600 At t = 700 At t = 800 At t = 900

P1 does a down P1 does a down P2 does an up P3 does a down P1 does a down P2 does an up P2 does a down P1 does an up P1 does an up

26. In an airline reservation system, it is necessary to ensure that while one process is busy using a file, no other process can also use it. Otherwise, two different processes, working for two different ticket agents, might each inadvertently sell the last seat on some flight. Devise a synchronization method using semaphores that makes sure that only one process at a time accesses each file (assuming that the processes obey the rules). 27. To make it possible to implement semaphores on a computer with multiple CPUs that share a common memory, computer architects often provide a Test and Set Lock instruction. TSL X tests the location X. If the contents are zero, they are set to 1 in a single, indivisible memory cycle, and the next instruction is skipped. If it is nonzero, the TSL acts like a no-op. Using TSL it is possible to write procedures lock and unlock with the following properties. lock (x) checks to see if x is locked. If not, it locks x and returns control. If x is already locked, it just waits until it becomes unlocked, then it locks x and returns control. unlock releases an existing lock. If all processes lock the semaphore table before using it, only one process at a time can fiddle with the variables and pointers, thus preventing races. Write lock and unlock in assembly language. (Make any reasonable assumptions you need.) 28. Show the values of in and out for a circular buffer of length 65 words after each of the following operations. Both start at 0. a. 22 words are put in b. 9 words are removed c. 40 words are put in d. 17 words are removed e. 12 words are put in f. 45 words are removed g. 8 words are put in h. 11 words are removed 29. Suppose that a version of UNIX uses 2-KB disk blocks and stores 512 disk addresses per indirect block (single, double, and triple). What would the maximum file size be? (Assume that file pointers are 64 bits wide).

CHAP. 6

PROBLEMS

505

30. Suppose that the UNIX system call unlink(′′/usr/ast/bin/game3′′)

were executed in the context of Figure 6-36. Describe carefully what changes are made in the directory system. 31. Imagine that you had to implement the UNIX system on a microcomputer where main memory was in short supply. After a considerable amount of shoehorning, it still did not quite fit, so you picked a system call at random to sacrifice for the general good. You picked pipe, which creates the pipes used to send byte streams from one process to another. Is it still possible to implement I/O redirection somehow? What about pipelines? Discuss the problems and possible solutions. 32. The Committee for Fairness to File Descriptors is organizing a protest against the UNIX system because whenever the latter returns a file descriptor, it always returns the lowest number not currently in use. Consequently, higher-numbered file descriptors are hardly ever used. Their plan is to return the lowest number not yet used by the program rather than the lowest number currently not in use. They claim that it is trivial to implement, will not affect existing programs, and is fairer. What do you think? 33. In XP it is possible to set up an access control list in such a way that Roberta has no access at all to a file, but everyone else has full access to it. How do you think this is implemented? 34. Describe two different ways to program producer-consumer problems using shared buffers and semaphores in XP. Think about how to implement the shared buffer in each case. 35. It is common to test out page replacement algorithms by simulation. For this exercise, you are to write a simulator for a page-based virtual memory for a machine with 64 1-KB pages. The simulator should maintain a single table of 64 entries, one per page, containing the physical page number corresponding to that virtual page. The simulator should read in a file containing virtual addresses in decimal, one address per line. If the corresponding page is memory, just record a page hit. If it is not in memory, call a page replacement procedure to pick a page to evict (i.e., an entry in the table to overwrite) and record a page miss. No page transport actually occurs. Generate a file consisting of random addresses and test the performance for both LRU and FIFO. Now generate an address file in which x percent of the addresses are four bytes higher than the previous one (to simulate locality). Run tests for various values of x and report on your results. 36. The program of Fig. 6-25 has a fatal race condition because two threads access shared variables in an uncontrolled way, without using semaphores or any other mutual exclusion technique. Run this program and see how long it takes to hang. If you cannot make it hang, modify it to increase the size of the window of vulnerability by putting some computing between adjusting m.in and m.out and testing them. How much computing do you have to put in before it fails, say, once an hour?

506

THE OPERATING SYSTEM MACHINE LEVEL

CHAP. 6

37. Write a program for UNIX or XP that takes as input the name of a directory. The program should print a list of the files in the directory, one line per file, and after the file name, print the size of the file. Print the file names in the order they occur in the directory. Unused slots in the directory should be listed as (unused).

7 THE ASSEMBLY LANGUAGE LEVEL

In Chapters 4, 5, and 6 we discussed three different levels present on most contemporary computers. This chapter is concerned primarily with another level that is also present on essentially all modern computers: the assembly language level. The assembly language level differs in a significant respect from the microarchitecture, ISA, and operating system machine levels—it is implemented by translation rather than by interpretation. Programs that convert a user’s program written in some language to another language are called translators. The language in which the original program is written is called the source language and the language to which it is converted is called the target language. Both the source language and the target language define levels. If a processor that can directly execute programs written in the source language is available, there is no need to translate the source program into the target language. Translation is used when a processor (either hardware or an interpreter) is available for the target language but not for the source language. If the translation has been performed correctly, running the translated program will give precisely the same results as the execution of the source program would have given had a processor for it been available. Consequently, it is possible to implement a new level for which there is no processor by first translating programs written for that level to a target level and then executing the resulting target-level programs. It is important to note the difference between translation, on the one hand, and interpretation, on the other hand. In translation, the original program in the source 507

508

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

language is not directly executed. Instead, it is converted to an equivalent program called an object program or executable binary program whose execution is carried out only after the translation has been completed. In translation, there are two distinct steps: 1. Generation of an equivalent program in the target language. 2. Execution of the newly generated program. These two steps do not occur simultaneously. The second step does not begin until the first has been completed. In interpretation, there is only one step: executing the original source program. No equivalent program need be generated first, although sometimes the source program is converted to an intermediate form (e.g., Java byte code) for easier interpretation. While the object program is being executed, only three levels are in evidence: the microarchitecture level, the ISA level, and the operating system machine level. Consequently, three programs—the user’s object program, the operating system, and the microprogram (if any)—can be found in the computer’s memory at run time. All traces of the original source program have vanished. Thus the number of levels present at execution time may differ from the number of levels present before translation. It should be noted, however, that although we define a level by the instructions and linguistic constructs available to its programmers (and not by the implementation technique), other authors sometimes make a greater distinction between levels implemented by execution-time interpreters and levels implemented by translation.

7.1 INTRODUCTION TO ASSEMBLY LANGUAGE Translators can be roughly divided into two groups, depending on the relation between the source language and the target language. When the source language is essentially a symbolic representation for a numerical machine language, the translator is called an assembler and the source language is called an assembly language. When the source language is a high-level language such as Java or C and the target language is either a numerical machine language or a symbolic representation for one, the translator is called a compiler.

7.1.1 What Is an Assembly Language? A pure assembly language is a language in which each statement produces exactly one machine instruction. In other words, there is a one-to-one correspondence between machine instructions and statements in the assembly program. If each line in the assembly language program contains exactly one statement and

SEC. 7.1

INTRODUCTION TO ASSEMBLY LANGUAGE

509

each machine word contains exactly one machine instruction, then an n-line assembly program will produce an n-word machine language program. The reason that people use assembly language, as opposed to programming in machine language (in hexadecimal), is that it is much easier to program in assembly language. The use of symbolic names and symbolic addresses instead of binary or octal ones makes an enormous difference. Most people can remember that the abbreviations for add, subtract, multiply, and divide are ADD, SUB, MUL, and DIV, but few can remember the corresponding numerical values the machine uses. The assembly language programmer need only remember the symbolic names because the assembler translates them to the machine instructions. The same remarks apply to addresses. The assembly language programmer can give symbolic names to memory locations and have the assembler worry about supplying the correct numerical values. The machine language programmer must always work with the numerical values of the addresses. As a consequence, no one programs in machine language today, although people did so decades ago, before assemblers had been invented. Assembly languages have another property, besides the one-to-one mapping of assembly language statements onto machine instructions, that distinguishes them from high-level languages. The assembly programmer has access to all the features and instructions available on the target machine. The high-level language programmer does not. For example, if the target machine has an overflow bit, an assembly language program can test it, but a Java program cannot directly test it. Such a program can execute every instruction in the instruction set of the target machine, but the high-level language program cannot. In short, everything that can be done in machine language can be done in assembly language, but many instructions, registers, and similar features are not available for the high-level language programmer to use. Languages for system programming, like C, are often a cross between these types, with the syntax of a high-level language but with much of the access to the machine of an assembly language. One final difference that is worth making explicit is that an assembly language program can run only on one family of machines, whereas a program written in a high-level language can potentially run on many machines. For many applications, this ability to move software from one machine to another is of great practical importance.

7.1.2 Why Use Assembly Language? Assembly language programming is difficult. Make no mistake about that. It is not for wimps and weaklings. Furthermore, writing a program in assembly language takes much longer than writing the same program in a high-level language. It also takes much longer to debug and is much harder to maintain. Under these conditions, why would anyone ever program in assembly language? There are two reasons: performance and access to the machine. First of

510

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

all, an expert assembly language programmer can often produce code that is much smaller and much faster than a high-level language programmer can. For some applications, speed and size are critical. Many embedded applications, such as the code on a smart card, the code in a cellular telephone, device drivers, BIOS routines, and the inner loops of performance-critical applications fall in this category. Second, some procedures need complete access to the hardware, something usually impossible in high-level languages. For example, the low-level interrupt and trap handlers in an operating system, and the device controllers in many embedded real-time systems fall into this category. The first reason for programming in assembly language (to get high performance) is usually the most important one, so let us look at that more closely. In most programs, a small percentage of the total code is responsible for a large percentage of the execution time. It is common to have 1% of the program be responsible for 50% of the execution time and 10% of the program be responsible for 90% of the execution time. Assume, for example, that it requires 10 programmer-years to write some program in a high-level language and that the resulting program requires 100 sec to execute a certain typical benchmark. (A benchmark is a test program used to compare computers, compilers, etc.) Writing the whole program in assembly language might require 50 programmer-years, due to the lower productivity of assembler language programmers. The final program might run the benchmark in about 33 sec, because a clever programmer can outdo a clever compiler by a factor of 3 (although you can get endless arguments about these ratios). This situation is illustrated in Fig. 7-1. Based on the above observation that only a tiny fraction of the code is responsible for most of the execution time, another approach is possible. The program is first written in a high-level language. Then a series of measurements is performed to determine which parts of the program account for most of the execution time. Such measurements would normally include using the system clock to compute the amount of time spent in each procedure, keeping track of the number of times each loop is executed, and similar steps. As an example, let us assume that 10% of the total program accounts for 90% of the execution time. This means that for a 100-sec job, 90 sec is spent in this critical 10% and 10 sec is spent in the remaining 90% of the program. The critical 10% can now be improved by rewriting it in assembly language. This process is called tuning and is illustrated in Fig. 7-1. Here an additional five programmeryears are needed to rewrite the critical procedures but their execution time is reduced from 90 sec to 30 sec. It is instructive to compare the mixed high-level language/assembly language approach with the pure assembly language version (see Fig. 7-1). The latter is about 20% faster (33 sec versus 40 sec) but at more than triple the price (50 programmer-years versus 15 programmer years). Furthermore, the advantage of

SEC. 7.1

INTRODUCTION TO ASSEMBLY LANGUAGE

511

######################################################################## ! Programmer-years to ! Program execution ! ! ! produce the program ! time in seconds ! # ####################################################################### ! ! ! ! 33 ! 50 ! ! Assembly language ! ! High-level language ! 10 ! 100 ! ! ! ! ! ! ! ! ! Mixed approach before tuning ! ! ! ! Critical 10% 1 ! 90 ! ! ! ! ! Other 90% 9 ! 10 ! ! ! "" ! "" ! ! ! Total 10 ! 100 ! ! ! ! ! ! ! ! ! ! Mixed approach after tuning ! ! ! ! ! Critical 10% 6 ! 30 ! ! ! Other 90% 9 ! 10 ! ! ! ! ! "" "" ! ! ! ! Total 15 ! 40 ! !######################################################################## ! !

Figure 7-1. Comparison of assembly language and high-level language programming, with and without tuning.

the mixed approach is really more than indicated, because recoding an already debugged high-level language procedure in assembly code is, in fact, much easier than writing the same assembly code procedure from scratch. In other words, the estimate of 5 programmer-years to rewrite the critical procedures is exceedingly conservative. If this recoding actually took only 1 programmer-year, the cost ratio between the mixed approach and the pure assembly language approach would be more than 4 to 1 in favor of the mixed approach. A programmer who uses a high-level language is not immersed in moving bits around and sometimes obtains insights into the problem that allow real improvements in performance. This situation rarely occurs with assembly language programmers, who are usually trying to juggle instructions to save a few cycles. All in all, there are still at least four good reasons for studying assembly language. First, because the success or failure of a large project may depend on being able to squeeze a factor of 2 or 3 improvement in performance out of some critical procedure, it is important to be able to write good assembly language code when it is really necessary. Second, assembly code is sometimes the only alternative due to lack of memory. Smart cards contain a CPU, but few have a megabyte of memory, and fewer yet have a hard disk for paging. Yet they must perform complex cryptographic calculations with limited resources. Processors embedded in appliances often have minimal memory for reasons of cost. Personal digital assistants and other battery-powered wireless electronic devices often have small memories to conserve battery power, so small, efficient code is a necessity here, too.

512

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

Third, a compiler must either produce output used by an assembler or perform the assembly process itself. Thus understanding assembly language is essential to understanding how compilers work. And someone has to write the compiler (and its assembler) after all. Finally, studying assembly language exposes the real machine to view. For students of computer architecture, writing some assembly code is the only way to get a feel for what machines are really like at the architectural level.

7.1.3 Format of an Assembly Language Statement Although the structure of an assembly language statement closely mirrors the structure of the machine instruction that it represents, assembly languages for different machines and different levels have sufficient resemblance to one another to allow a discussion of assembly language in general. Figure 7-2 shows fragments of assembly language programs for the Pentium 4, Motorola 680x0, and the (Ultra)SPARC which perform the computation N = I + J. In all three examples, the statements above the blank line perform the computation. The statements below the blank line are commands to the assembler to reserve memory for the variables I, J, and N and are not symbolic representations of machine instructions. Several assemblers exist for the Intel family, each with a different syntax. In this chapter we will use the Microsoft MASM assembly language for our examples. Although we are focusing on the Pentium 4, all of what we have to say about it applies equally well to the 386, 486, Pentium, and Pentium Pro. For the SPARC, we are basing our examples on the Sun assembler. Here, too, everything we say applies to the earlier (32-bit) versions of the SPARC as well. For the sake of uniformity, we will use uppercase for opcodes and registers throughout (Pentium 4 convention) even though the Sun assembler expects lowercase. Assembly language statements have four parts: a label field, an operation (opcode) field, an operands field, and a comments field. Labels, which are used to provide symbolic names for memory addresses, are needed on executable statements so that the statements can be branched to. They are also needed for data words to permit the data stored there to be accessible by symbolic name. If a statement is labeled, the label (usually) begins in column 1. Each of the three parts of Fig. 7-2 has four labels: FORMULA, I, J, and N. Notice that the SPARC assembly languages requires a colon after each label, but the Motorola one does not. The Intel one requires colons on code labels but not on data labels. There is nothing fundamental about this difference. The designers of different assemblers often have different tastes. Nothing in the underlying architecture suggests one choice or the other. One advantage of the colon notation is that with it a label can appear by itself on a line, with the opcode in column 1 of the next line. This style is sometimes convenient for compilers. Without the colon, there would be no way to tell a label on a line all by itself from an opcode on a line all by itself. The colon eliminates this potential ambiguity.

SEC. 7.1

INTRODUCTION TO ASSEMBLY LANGUAGE

513

Comments Operands Opcode Label ################################################################################# FORMULA: MOV EAX,I ; register EAX = I ; register EAX = I + J EAX,J ADD MOV N,EAX ;N=I+J I J N

DD DD DD

; reserve 4 bytes initialized to 3 ; reserve 4 bytes initialized to 4 ; reserve 4 bytes initialized to 0

3 4 0 (a)

Label Opcode Operands Comments ################################################################################ FORMULA MOVE.L I, D0 ; register D0 = I ADD.L J, D0 ; register D0 = I + J MOVE.L D0, N ;N=I+J I J N

DC.L DC.L DC.L

3 4 0

(b)

; reserve 4 bytes initialized to 3 ; reserve 4 bytes initialized to 4 ; reserve 4 bytes initialized to 0

Opcode Operands Comments #Label ################################################################################### FORMULA: SETHI %HI(I),%R1 ! R1 = high-order bits of the address of I LD [%R1+%LO(I)],%R1 ! R1 = I SETHI %HI(J),%R2 ! R2 = high-order bits of the address of J LD [%R2+%LO(J)],%R2 ! R2 = J NOP ! wait for J to arrive from memory ADD %R1,%R2,%R2 ! R2 = R1 + R2 SETHI %HI(N),%R1 ! R1 = high-order bits of the address of N ST %R2,[%R1+%LO(N)] I: J: N:

.WORD 3 .WORD 4 .WORD 0

(c)

! reserve 4 bytes initialized to 3 ! reserve 4 bytes initialized to 4 ! reserve 4 bytes initialized to 0

Figure 7-2. Computation of N = I + J. (a) Pentium 4. (b) Motorola 680x0. (c) SPARC.

It is an unfortunate characteristic of some assemblers that labels are restricted to six or eight characters. In contrast, most high-level languages allow the use of arbitrarily long names. Long, well-chosen names make programs much more readable and understandable by other people. Each of the machines has some registers, but they have been given very different names. The Pentium 4 registers have names like EAX, EBX, ECX, and so on.

514

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

The Motorola registers are called D0, D1, D2, among others. The SPARC registers have multiple names. Here we have used %R1 and %R2 for them. The opcode field contains either a symbolic abbreviation for the opcode—if the statement is a symbolic representation for a machine instruction—or a command to the assembler itself. The choice of an appropriate name is just a matter of taste, and different assembly language designers often make different choices. The designers of the Intel assembler decided to use MOV for both loading a register from memory and storing a register into memory. The designers of the Motorola assembler chose MOVE for both of them. In contrast, the designers of the SPARC assembler decided to use LD for the former and ST for the latter. Here, too, these choices have nothing to do with the underlying machine. In contrast, the need to use two machine instructions, starting with SETHI, to access memory, is an inherent property of the SPARC architecture because virtual addresses are 32 bits (SPARC Version 8) or 44 (SPARC Version 9) bits and instructions can hold at most 22 bits of immediate data. Thus it always takes two instructions to provide all the bits of a full virtual address. What SETHI %HI(I),%R1

does is zero the upper 32 bits and lower 10 bits of (the 64-bit) register R1, then put the upper 22 bits of the 32-bit address of I in bit positions 10 through 31 of R1. The next instruction, LD [%R1+%LO(I)],%R1

adds R1 and the low-order 10 bits of the address of I to form the full address of I, fetch that word from memory, and put it in R1. In a beauty contest using a scale of 1 to 10, these instructions would score about −20, but the SPARC was not designed for the beauty of its assembly language. It was designed for high-speed execution, and it achieves that goal well. The Pentium family, 680x0, and SPARC all allow byte, word, and long operands. How does the assembler know which length to use? Again, the assembler designers chose different solutions. On the Pentium 4, different length registers have different names, so EAX is used to move 32-bit items, AX is used to move 16-bit items, and AL and AH are used to move 8-bit items. The Motorola assembler designers, in contrast decided to add a suffix .L for long, .W for word, or .B for byte to each opcode rather than giving subsets of D0, etc., different names. The SPARC uses different opcodes for the different lengths (e.g., LDSB, LDSH, and LDSW to load signed bytes, halfwords, and words into a 64-bit register, respectively). All three ways are valid, but again they point out the arbitrary nature of language design. The three assemblers also differ in how they reserve space for data. The Intel assembly language designers chose DD (Define Double), since a word on the 8088 was 16 bits. The Motorola ones liked DC (Define Constant). The SPARC folks preferred .WORD from the beginning. Once again, the differences are arbitrary.

SEC. 7.1

INTRODUCTION TO ASSEMBLY LANGUAGE

515

The operands field of an assembly language statement is used to specify the addresses and registers used as operands by the machine instruction. The operands field of an integer addition instruction tells what is to be added to what. The operands field of a branch instruction tells where to branch to. Operands can be registers, constants, memory locations, and so on. The comments field provides a place for programmers to put helpful explanations of how the program works for the benefit of other programmers who may subsequently use or modify the program (or for the benefit of the original programmer a year later). An assembly language program without such documentation is nearly incomprehensible to all programmers, frequently including the author as well. The comments field is solely for human consumption; it has no effect on the assembly process or on the generated program.

7.1.4 Pseudoinstructions In addition to specifying which machine instructions to execute, an assembly language program can also contain commands to the assembler itself, for example, asking it to allocate some storage or to eject to a new page on the listing. Commands to the assembler itself are called pseudoinstructions or sometimes assembler directives. We have already seen a typical pseudoinstruction in Fig. 7-2(a): DW. Some other pseudoinstructions are listed in Fig. 7-3. These are taken from the Microsoft MASM assembler for the Intel family. The SEGMENT pseudoinstruction starts a new segment, and ENDS terminates one. It is allowed to start a text segment, with code, then start a data segment, then go back to the text segment, and so on. ALIGN forces the next line, usually data, to an address that is a multiple of its argument. For example, if the current segment has 61 bytes of data already, then after ALIGN 4 the next address allocated will be 64. EQU is used to give a symbolic name to an expression. For example, after the pseudoinstruction BASE EQU 1000

the symbol BASE can be used everywhere instead of 1000. The expression that follows the EQU can involve multiple defined symbols combined with arithmetic and other operators, as in LIMIT EQU 4 * BASE + 2000

Most assemblers, including MASM, require that a symbol be defined before it is used in an expression like this. The next four pseudoinstructions, DB, DW, DD, and DQ, allocate storage for one or more variables of size 1, 2, 4, or 8 bytes, respectively. For example, TABLE DB 11, 23, 49

516

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

##################################################################################### ! Pseudoinstruction ! ! Meaning ##################################################################################### ! ! ! ! SEGMENT ! Start a new segment (text, data, etc.) with certain attributes ! ##################################################################################### ! ENDS ! End the current segment ! ##################################################################################### ! ! ! ! ALIGN ! Control the alignment of the next instruction or data ! ##################################################################################### ! ! ! EQU Define a new symbol equal to a given expression ##################################################################################### ! ! ! ! DB ! Allocate storage for one or more (initialized) bytes ! ##################################################################################### ! ! ! ! DW ! Allocate storage for one or more (initialized) 16-bit (word) data items ! ##################################################################################### ! DD ! Allocate storage for one or more (initialized) 32-bit (double) data items ! ##################################################################################### ! ! ! ! DQ ! Allocate storage for one or more (initialized) 64-bit (quad) data items ! ##################################################################################### ! PROC ! Start a procedure ! ##################################################################################### ! ! ! ! ENDP ! End a procedure ! ##################################################################################### ! ! ! ##################################################################################### ! MACRO ! Start a macro definition ! ! ENDM ! End a macro definition ! ##################################################################################### ! ! ! ! PUBLIC ! Export a name defined in this module ! ##################################################################################### ! EXTERN ! Import a name from another module ! ##################################################################################### ! ! ! ! INCLUDE ! Fetch and include another file ! ##################################################################################### ! ! ! ##################################################################################### ! IF ! Start conditional assembly based on a given expression ! ! ELSE ! Start conditional assembly if the IF condition above was false ! ##################################################################################### ! ! ! ##################################################################################### ! ENDIF ! End conditional assembly ! ! COMMENT ! Define a new start-of-comment character ! ##################################################################################### ! ! ! ! PAGE ! Generate a page break in the listing ! ##################################################################################### ! END ! Terminate the assembly program ! ! ##################################################################################### ! ! Figure 7-3. Some of the pseudoinstructions available in the Pentium 4 assembler (MASM).

allocates space for 3 bytes and initializes them to 11, 23, and 49, respectively. It also defines the symbol TABLE and sets it equal to the address where 11 is stored. The PROC and ENDP pseudoinstructions define the start and end of assembly language procedures, respectively. Procedures in assembly language have the same function as procedures in other programming languages. Similarly, MACRO and ENDM delimit the scope of a macro definition. We will study macros later in this chapter. The next two pseudoinstructions, PUBLIC and EXTERN, control the visibility of symbols. It is common to write programs as a collection of files. Frequently, a procedure in one file needs to call a procedure or access a data word defined in another file. To make this cross-file referencing possible, a symbol that is to be made available to other files is exported using PUBLIC. Similarly, to prevent the assembler from complaining about the use of a symbol that is not defined in the current file, the symbol can be declared as EXTERN, which tells the assembler that

SEC. 7.1

INTRODUCTION TO ASSEMBLY LANGUAGE

517

it will be defined in some other file. Symbols that are not declared in either of these pseudoinstructions have a scope of the local file. This default means that using, say, FOO in multiple files does not generate a conflict because each definition is local to its own file. The INCLUDE pseudoinstruction causes the assembler to fetch another file and include it bodily into the current one. Such included files often contain definitions, macros, and other items needed in multiple files. Many assemblers, including MASM, support conditional assembly. For example, WORDSIZE EQU 16 IF WORDSIZE GT 16 WSIZE: DD 32 ELSE WSIZE: DD 16 ENDIF

allocates a single 32-bit word and calls its address WSIZE. The word is initialized to either 32 or 16, depending on the value of WORDSIZE, in this case, 16. Typically this construction would be used to write a program that could be assembled on either 16-bit machines (like the 8088) or 32-bit machines (like the Pentium 4). By bracketing all the machine-dependent code within IF and ENDIF, then by changing a single definition, WORDSIZE, the program can automatically be set to assemble for either size. Using this approach, it is possible to maintain one source program for multiple (different) target machines, which makes software development and maintenance easier. In many cases, all the machine-dependent definitions, like WORDSIZE, are collected into a single file, with different versions for different machines. By including the right definitions file, the program can be easily recompiled for different machines. The COMMENT pseudoinstruction allows the user to change the comment delimiter to something other than semicolon. PAGE is used to control the listing the assembler can produce, if requested. Finally, END marks the end of the program. Many other pseudoinstructions exist in MASM. Other Pentium 4 assemblers have a different collection of pseudoinstructions available because they are dictated not by the underlying architecture, but by the taste of the assembler writer.

7.2 MACROS Assembly language programmers frequently need to repeat sequences of instructions several times within a program. The most obvious way to do so is simply to write the required instructions wherever they are needed. If a sequence is long, however, or must be used a large number of times, writing it repeatedly becomes tedious. An alternative approach is to make the sequence into a procedure and call it wherever it is needed. This strategy has the disadvantage of requiring a procedure

518

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

call instruction and a return instruction to be executed every time a sequence is needed. If the sequences are short—for example, two instructions—but are used frequently, the procedure call overhead may significantly slow the program down. Macros provide an easy and efficient solution to the problem of repeatedly needing the same or nearly the same sequences of instructions.

7.2.1 Macro Definition, Call, and Expansion A macro definition is a way to give a name to a piece of text. After a macro has been defined, the programmer can write the macro name instead of the piece of program. A macro is, in effect, an abbreviation for a piece of text. Figure 74(a) shows an assembly language program for the Pentium 4 that exchanges the contents of the variables p and q twice. These sequences could be defined as macros, as shown in Fig. 7-4(b). After its definition, every occurrence of SWAP causes it to be replaced by the four lines: MOV EAX,P MOV EBX,Q MOV Q,EAX MOV P,EBX

The programmer has defined SWAP as an abbreviation for the four statements shown above. MOV MOV MOV MOV

EAX,P EBX,Q Q,EAX P,EBX

MOV MOV MOV MOV

EAX,P EBX,Q Q,EAX P,EBX

SWAP

(a)

MACRO MOV EAX,P MOV EBX,Q MOV Q,EAX MOV P,EBX ENDM SWAP SWAP (b)

Figure 7-4. Assembly language code for interchanging P and Q twice. (a) Without a macro. (b) With a macro.

Although different assemblers have slightly different notations for defining macros, all require the same basic parts in a macro definition: 1. A macro header giving the name of the macro being defined. 2. The text comprising the body of the macro. 3. A pseudoinstruction marking the end of the definition (e.g., ENDM).

SEC. 7.2

MACROS

519

When the assembler encounters a macro definition, it saves it in a macro definition table for subsequent use. From that point on, whenever the name of the macro (SWAP in the example of Fig. 7-4) appears as an opcode, the assembler replaces it by the macro body. The use of a macro name as an opcode is known as a macro call and its replacement by the macro body is called macro expansion. Macro expansion occurs during the assembly process and not during execution of the program. This point is important. The program of Fig. 7-4(a) and that of Fig. 7-4(b) will produce precisely the same machine language code. Looking only at the machine language program, it is impossible to tell whether or not any macros were involved in its generation. The reason is that once macro expansion has been completed the macro definitions are discarded by the assembler. No trace of them is left in the generated program. Macro calls should not be confused with procedure calls. The basic difference is that a macro call is an instruction to the assembler to replace the macro name with the macro body. A procedure call is a machine instruction that is inserted into the object program and that will later be executed to call the procedure. Figure 7-5 compares macro calls with procedure calls. #################################################################################### ! ! ! ! Item Macro call Procedure call #! ################################################################################### ! ! ! When is the call made? ! During assembly ! During program execution ! #! ################################################################################### ! Is the body inserted into the object ! Yes ! No ! ! ! ! ! ! program every place the call is ! ! ! ! ! ! made? #! ################################################################################### ! ! ! ! ! Is a procedure call instruction ! No ! Yes ! ! inserted into the object program ! ! ! ! ! ! and later executed? #! ################################################################################### ! ! ! ! ! Must a return instruction be used ! No ! Yes ! ! ! ! after the call is done? #! ################################################################################### ! ! ! ! ! How many copies of the body ap! One per macro call ! One ! pear in the object program? !#################################################################################### ! ! ! Figure 7-5. Comparison of macro calls with procedure calls.

Conceptually, it is best to think of the assembly process as taking place in two passes. On pass one, all the macro definitions are saved and the macro calls expanded. On pass two, the resulting text is processed as though it was in the original program. In this view, the source program is read in and is then transformed into another program from which all macro definitions have been removed, and in which all macro calls have been replaced by their bodies. The resulting output, an assembly language program containing no macros at all, is then fed into the assembler. It is important to keep in mind that a program is a string of characters including letters, digits, spaces, punctuation marks, and ‘‘carriage returns’’ (change to a

520

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

new line). Macro expansion consists of replacing certain substrings of this string with other character strings. A macro facility is a technique for manipulating character strings, without regard to their meaning.

7.2.2 Macros with Parameters The macro facility previously described can be used to shorten programs in which precisely the same sequence of instructions occurs repeatedly. Frequently, however, a program contains several sequences of instructions that are almost but not quite identical, as illustrated in Fig. 7-6(a). Here the first sequence exchanges P and Q, and the second sequence exchanges R and S. MOV MOV MOV MOV

EAX,P EBX,Q Q,EAX P,EBX

MOV MOV MOV MOV

EAX,R EBX,S S,EAX R,EBX

(a)

CHANGE

MACRO P1, P2 MOV EAX,P1 MOV EBX,P2 MOV P2,EAX MOV P1,EBX ENDM CHANGE P, Q CHANGE R, S (b)

Figure 7-6. Nearly identical sequences of statements. (a) Without a macro. (b) With a macro.

Macro assemblers handle the case of nearly identical sequences by allowing macro definitions to provide formal parameters and by allowing macro calls to supply actual parameters. When a macro is expanded, each formal parameter appearing in the macro body is replaced by the corresponding actual parameter. The actual parameters are placed in the operand field of the macro call. Figure 7-6(b) shows the program of Fig. 7-6(a) rewritten using a macro with two parameters. The symbols P1 and P2 are the formal parameters. Each occurrence of P1 within a macro body is replaced by the first actual parameter when the macro is expanded. Similarly, P2 is replaced by the second actual parameter. In the macro call CHANGE P, Q

P is the first actual parameter and Q is the second actual parameter. Thus the executable programs produced by both parts of Fig. 7-6 are identical. They contain precisely the same instructions with the same operands.

SEC. 7.2

MACROS

521

7.2.3 Advanced Features Most macro processors have a whole raft of advanced features to make life easier for the assembly language programmer. In this section we will take a look at a few of MASM’s advanced features. One problem that occurs with all assemblers that support macros is label duplication. Suppose that a macro contains a conditional branch instruction and a label that is branched to. If the macro is called two or more times, the label will be duplicated, causing an assembly error. One solution is to have the programmer supply a different label on each call as a parameter. A different solution (used by MASM) is to allow a label to be declared LOCAL, with the assembler automatically generating a different label on each expansion of the macro. Some other assemblers have a rule that numeric labels are automatically local. MASM and most other assemblers allow macros to be defined within other macros. This feature is most useful in combination with conditional assembly. Typically, the same macro is defined in both parts of an IF statement, like this: M1 M2

M2

MACRO IF WORDSIZE GT 16 MACRO ... ENDM ELSE MACRO ... ENDM ENDIF ENDM

Either way, the macro M2 will be defined, but the definition will depend on whether the program is being assembled on a 16-bit machine or a 32-bit machine. If M1 is not called, M2 will not be defined at all. Finally, macros can call other macros, including themselves. If a macro is recursive, that is, it calls itself, it must pass itself a parameter that is changed on each expansion and the macro must test the parameter and terminate the recursion when it reaches a certain value. Otherwise the assembler can be put into an infinite loop. If this happens, the assembler must be killed explicitly by the user.

7.2.4 Implementation of a Macro Facility in an Assembler To implement a macro facility, an assembler must be able to perform two functions: save macro definitions and expand macro calls. We will examine these functions in turn. The assembler must maintain a table of all macro names and, along with each name, a pointer to its stored definition so that it can be retrieved when needed.

522

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

Some assemblers have a separate table for macro names and some have a combined opcode table in which all machine instructions, pseudoinstructions, and macro names are kept. When a macro definition is encountered, a table entry is made giving the name of the macro, the number of formal parameters, and a pointer to another table—the macro definition table—where the macro body will be kept. A list of the formal parameters is also constructed at this time for use in processing the definition. The macro body is then read and stored in the macro definition table. Formal parameters occurring within the body are indicated by some special symbol. As an example, the internal representation of the macro definition of CHANGE with semicolon as ‘‘carriage return’’ and ampersand as the formal parameter symbol is shown below: MOV EAX,&P1; MOV EBX,&P2; MOV &P2,EAX; MOV &P1,EBX;

Within the macro definition table the macro body is simply a character string. During pass one of the assembly, opcodes are looked up and macros expanded. Whenever a macro definition is encountered, it is stored in the macro table. When a macro is called, the assembler temporarily stops reading input from the input device and starts reading from the stored macro body instead. Formal parameters extracted from the stored macro body are replaced by the actual parameters provided in the call. The presence of an ampersand in front of the formal parameters makes it easy for the assembler to recognize them.

7.3 THE ASSEMBLY PROCESS In the following sections we will briefly describe how an assembler works. Although each machine has a different assembly language, the assembly process is sufficiently similar on different machines that it is possible to describe it in general terms.

7.3.1 Two-Pass Assemblers Because an assembly language program consists of a series of one-line statements, it might at first seem natural to have an assembler that read one statement, then translated it to machine language, and finally output the generated machine language onto a file, along with the corresponding piece of the listing, if any, onto another file. This process would then be repeated until the whole program had been translated. Unfortunately, this strategy does not work. Consider the situation where the first statement is a branch to L. The assembler cannot assemble this statement until it knows the address of statement L. Statement L may be near the end of the program, making it impossible for the assembler to find the address without first reading almost the entire program.

SEC. 7.3

THE ASSEMBLY PROCESS

523

This difficulty is called the forward reference problem, because a symbol, L, has been used before it has been defined; that is, a reference has been made to a symbol whose definition will only occur later. Forward references can be handled in two ways. First, the assembler may in fact read the source program twice. Each reading of the source program is called a pass; any translator that reads the input program twice is called a two-pass translator. On pass one of a two-pass assembler, the definitions of symbols, including statement labels, are collected and stored in a table. By the time the second pass begins, the values of all symbols are known; thus no forward reference remains and each statement can be read, assembled, and output. Although this approach requires an extra pass over the input, it is conceptually simple. The second approach consists of reading the assembly program once, converting it to an intermediate form, and storing this intermediate form in a table in memory. Then a second pass is made over the table instead of over the source program. If there is enough memory (or virtual memory), this approach saves I/O time. If a listing is to be produced, then the entire source statement, including all the comments, has to be saved. If no listing is needed, then the intermediate form can be reduced to the bare essentials. Either way, another task of pass one is to save all macro definitions and expand the calls as they are encountered. Thus defining the symbols and expanding the macros are generally combined into one pass.

7.3.2 Pass One The principal function of pass one is to build up a table called the symbol table, containing the values of all symbols. A symbol is either a label or a value that is assigned a symbolic name by means of a pseudoinstruction such as BUFSIZE EQU 8192

In assigning a value to a symbol in the label field of an instruction, the assembler must know what address that instruction will have during execution of the program. To keep track of the execution-time address of the instruction being assembled, the assembler maintains a variable during assembly, known as the ILC (Instruction Location Counter). This variable is set to 0 at the beginning of pass one and incremented by the instruction length for each instruction processed, as shown in Fig. 7-7. This example is for the Pentium 4. We will not give SPARC (or Motorola) examples henceforth since the differences between the assembly languages are not very important, and one example should be enough. Besides, if there a contest for the world’s least readable assembly language, the SPARC’s would be a real contender. Pass one of most assemblers uses at least three internal tables: the symbol table, the pseudoinstruction table, and the opcode table. If needed, a literal table is also kept. The symbol table has one entry for each symbol, as illustrated in

524

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

Label Opcode Operands Comments Length ILC ############################################################################ MARIA: MOV EAX, I EAX = I 5 100 MOV EBX, J EBX = J 6 105 ROBERTA: MOV ECX, K ECX = K 6 111 2 117 IMUL EAX, EAX EAX = I * I 3 119 IMUL EBX, EBX EBX = J * J 3 122 IMUL ECX, ECX ECX = K * K 2 125 MARILYN: ADD EAX, EBX EAX = I * I + J * J 2 127 ADD EAX, ECX EAX = I * I + J * J + K * K STEPHANY: JMP DONE branch to DONE 5 129 Figure 7-7. The instruction location counter (ILC) keeps track of the address where the instructions will be loaded in memory. In this example, the statements prior to MARIA occupy 100 bytes.

Fig. 7-8. Symbols are defined either by using them as labels or by explicit definition (e.g., EQU). Each symbol table entry contains the symbol itself (or a pointer to it), its numerical value, and sometimes other information. This additional information may include 1. The length of data field associated with symbol. 2. The relocation bits. (Does the symbol change value if the program is loaded at a different address than the assembler assumed?) 3. Whether or not the symbol is to be accessible outside the procedure. ################################################# ! Symbol ! Value ! ! Other information #! ################################################ ! ! ! MARIA !################################################# ! 100 ! ! ! ROBERTA ! 111 ! ! #! ################################################ ! ! ! !################################################# ! 125 ! ! MARILYN ! ! ! ! STEPHANY ! 129 ! !################################################# ! Figure 7-8. A symbol table for the program of Fig. 7-7.

The opcode table contains at least one entry for each symbolic opcode (mnemonic) in the assembly language. Figure 7-9 shows part of an opcode table. Each entry contains the symbolic opcode, two operands, the opcode’s numerical value, the instruction length, and a type number that separates the opcodes into groups depending on the number and kind of operands. As an example, consider the opcode ADD. If an ADD instruction contains EAX as the first operand and a 32-bit constant (immed32) as the second one, then opcode 0x05 is used and the instruction length is 5 bytes. (Constants that can be expressed in 8 or 16 bits use different opcodes, not shown.) If ADD is used with two registers as operands, the instruction is 2 bytes, with opcode 0x01. The

SEC. 7.3

THE ASSEMBLY PROCESS

525

############################################################################ ! ! ! Second ! Hexadecimal ! Instruction ! Instruction ! First ! ! ! ! ! ! ! Opcode ! operand ! operand ! opcode length class !############################################################################ ! ! ! !############################################################################ ! ! ! ! ! ! AAA — — 37 1 6 ! ! ! ! ! ! ! ADD 05 5 4 !############################################################################ ! EAX ! immed32 ! ! ! ! ! ADD ! reg ! reg ! ! ! ! 01 2 19 # ########################################################################### ! ! ! ! ! ! ! AND 25 5 4 !############################################################################ ! EAX ! immed32 ! ! ! ! ! AND ! reg ! reg ! ! ! ! 21 2 19 !############################################################################ ! ! ! ! ! ! Figure 7-9. A few excerpts from the opcode table for a Pentium 4 assembler.

(arbitrary) instruction class 19 would be given to all opcode-operand combinations that follow the same rules and should be processed the same way as ADD with two register operands. The instruction class effectively designates a procedure within the assembler that is called to process all instructions of a given type. Some assemblers allow programmers to write instructions using immediate addressing even though no corresponding target language instruction exists. Such ‘‘pseudoimmediate’’ instructions are handled as follows. The assembler allocates memory for the immediate operand at the end of the program and generates an instruction that references it. For instance, the IBM 3090 mainframe has no immediate instructions. Nevertheless, programmers may write L 14,=F′5′

to load register 14 with a full word constant 5. In this way, the programmer avoids explicitly writing a pseudoinstruction to allocate a word initialized to 5, giving it a label, and then using that label in the L instruction. Constants for which the assembler automatically reserves memory are called literals. In addition to saving the programmer a little writing, literals improve the readability of a program by making the value of the constant apparent in the source statement. Pass one of the assembler must build a table of all literals used in the program. All three of our example computers have immediate instructions, so their assemblers do not provide literals. Immediate instructions are quite common nowadays, but formerly they were unusual. It is likely that the widespread use of literals made it clear to machine designers that immediate addressing was a good idea. If literals are needed, a literal table is maintained during assembly, with a new entry made each time a literal is encountered. After the first pass, this table is sorted and duplicates removed. Figure 7-10 shows a procedure that could serve as a basis for pass one of an assembler. The style of programming is noteworthy in itself. The procedure names have been chosen to give a good indication of what the procedures do. Most important, Fig. 7-10 represents an outline of pass one which, although not complete, forms a good starting point. It is short enough to be easily understood

526

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

and it makes clear what the next step must be—namely, to write the procedures used in it. public static void pass"one( ) { // This procedure is an outline of pass one of a simple assembler. boolean more"input = true; // flag that stops pass one String line, symbol, literal, opcode; // fields of the instruction int location"counter, length, value, type; // misc. variables final int END"STATEMENT = −2; // signals end of input location"counter = 0; initialize "tables( );

// assemble first instruction at 0 // general initialization

while (more "input) { line = read"next"line( ); length = 0; type = 0;

// more"input set to false by END // get a line of input // # bytes in the instruction // which type (format) is the instruction

if (line "is"not"comment(line)) { symbol = check"for"symbol(line); // is this line labeled? if (symbol != null) // if it is, record symbol and value enter"new"symbol(symbol, location"counter); literal = check"for"literal(line); // does line contain a literal? // if it does, enter it in table if (literal != null) enter"new"literal(literal);

}

}

}

// Now determine the opcode type. −1 means illegal opcode. opcode = extract"opcode(line); // locate opcode mnemonic type = search"opcode"table(opcode); // find format, e.g. OP REG1,REG2 if (type < 0) // if not an opcode, is it a pseudoinstruction? type = search"pseudo"table(opcode); // determine the length of this instruction switch(type) { case 1: length = get"length"of"type1(line); break; case 2: length = get"length"of"type2(line); break; // other cases here }

write "temp"file(type, opcode, length, line); // useful info for pass two location"counter = location"counter + length; // update loc"ctr if (type == END"STATEMENT) { // are we done with input? more"input = false; // if so, perform housekeeping tasks rewind "temp"for"pass"two( ); // like rewinding the temp file sort"literal "table( ); // and sorting the literal table remove"redundant"literals( ); // and removing duplicates from it }

Figure 7-10. Pass one of a simple assembler.

Some of these procedures will be relatively short, such as check"for"symbol, which just returns the symbol as a character string if there is one and null if there

SEC. 7.3

THE ASSEMBLY PROCESS

527

is not. Other procedures, such as get"length"of"type1 and get"length"of"type2, may be longer and may call other procedures. In general, the number of types will not be two, of course, but will depend on the language being assembled and how many types of instructions it has. Structuring programs in this way has other advantages in addition to ease of programming. If the assembler is being written by a group of people, the various procedures can be parceled out among the programmers. All the (nasty) details of getting the input are hidden away in read"next"line. If they should change—for example, due to an operating system change—only one subsidiary procedure is affected, and no changes are needed to the pass"one procedure itself. As it reads the program, pass one has to parse each line to find the opcode (e.g., ADD), look up its type (basically, the pattern of operands), and compute the instruction’s length. This information is also needed on the second pass, so it is possible to write it out explicitly to eliminate the need to parse the line from scratch next time. However, rewriting the input file causes more I/O to occur. Whether it is better to do more I/O to eliminate parsing or less I/O and more parsing depends on the relative speed of the CPU and disk, the efficiency of the file system, and other factors. In this example we will write out a temporary file containing the type, opcode, length, and actual input line. It is this line that pass two reads instead of the raw input file. When the END pseudoinstruction is read, pass one is over. The symbol table and literal tables can be sorted at this point if needed. The sorted literal table can be checked for duplicate entries, which can be removed.

7.3.3 Pass Two The function of pass two is to generate the object program and possibly print the assembly listing. In addition, pass two must output certain information needed by the linker for linking up procedures assembled at different times into a single executable file. Figure 7-11 shows a sketch of a procedure for pass two. The operation of pass two is more-or-less similar to that of pass one: it reads the lines one at a time and processes them one at a time. Since we have written the type, opcode, and length at the start of each line (on the temporary file), all of these are read in to save some parsing. The main work of the code generation is done by the procedures eval"type1, eval"type2, and so on. Each one handles a particular pattern, such as an opcode and two register operands. It generates the binary code for the instruction and returns it in code. Then it is written out. More likely, write"output just buffers the accumulated binary code and writes the file to disk in large chunks to reduce disk traffic. The original source statement and the object code generated from it (in hexadecimal) can then be printed or put into a buffer for later printing. After the ILC has been adjusted, the next statement is fetched.

528

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

public static void pass"two( ) { // This procedure is an outline of pass two of a simple assembler. // flag that stops pass two boolean more"input = true; String line, opcode; // fields of the instruction int location"counter, length, type; // misc. variables final int END"STATEMENT = −2; // signals end of input final int MAX"CODE = 16; // max bytes of code per instruction byte code[ ] = new byte[MAX"CODE]; // holds generated code per instruction location"counter = 0;

// assemble first instruction at 0

while (more "input) { type = read"type( ); opcode = read"opcode( ); length = read"length( ); line = read"line( );

// more"input set to false by END // get type field of next line // get opcode field of next line // get length field of next line // get the actual line of input

if (type != 0) { // type 0 is for comment lines switch(type) { // generate the output code case 1: eval"type1(opcode, length, line, code); break; case 2: eval"type2(opcode, length, line, code); break; // other cases here } }

}

}

write "output(code); // write the binary code write "listing(code, line); // print one line on the listing location"counter = location"counter + length; // update loc"ctr if (type == END"STATEMENT) { // are we done with input? more"input = false; // if so, perform housekeeping tasks finish"up( ); // odds and ends }

Figure 7-11. Pass two of a simple assembler.

Up until now it has been assumed that the source program does not contain any errors. Anyone who has ever written a program, in any language, knows how realistic that assumption is. Some of the common errors are as follows: 1. A symbol has been used but not defined. 2. A symbol has been defined more than once. 3. The name in the opcode field is not a legal opcode. 4. 5. 6. 7. 8.

An opcode is not supplied with enough operands. An opcode is supplied with too many operands. An octal number contains an 8 or a 9. Illegal register use (e.g., a branch to a register). The END statement is missing.

SEC. 7.3

THE ASSEMBLY PROCESS

529

Programmers are most ingenious at thinking up new kinds of errors to make. Undefined symbol errors are frequently caused by typing errors, so a clever assembler could try to figure out which of the defined symbols most resembles the undefined one and use that instead. Little can be done about correcting most other errors. The best thing for the assembler to do with an errant statement is to print an error message and try to continue assembly.

7.3.4 The Symbol Table During pass one of the assembly process, the assembler accumulates information about symbols and their values that must be stored in the symbol table for lookup during pass two. Several different ways are available for organizing the symbol table. We will briefly describe some of them below. All of them attempt to simulate an associative memory, which conceptually is a set of (symbol, value) pairs. Given the symbol, the associative memory must produce the value. The simplest implementation technique is indeed to implement the symbol table as an array of pairs, the first element of which is (or points to) the symbol and the second of which is (or points to) the value. Given a symbol to look up, the symbol table routine just searches the table linearly until it finds a match. This method is easy to program but is slow, because, on the average, half the table will have to be searched on each lookup. Another way to organize the symbol table is to sort it on the symbols and use the binary search algorithm to look up a symbol. This algorithm works by comparing the middle entry in the table to the symbol. If the symbol comes before the middle entry alphabetically, the symbol must be located in the first half of the table. If the symbol comes after the middle entry, it must be in the second half of the table. If the symbol is equal to the middle entry, the search terminates. Assuming that the middle entry is not equal to the symbol sought, we at least know which half of the table to look for it in. Binary search can now be applied to the correct half, which yields either a match, or the correct quarter of the table. Applying the algorithm recursively, a table of size n entries can be searched in about log2 n attempts. Obviously, this way is much faster than searching linearly, but it requires maintaining the table in sorted order. A completely different way of simulating an associative memory is a technique known as hash coding or hashing. This approach requires having a ‘‘hash’’ function that maps symbols onto integers in the range 0 to k − 1. One possible function is to multiply the ASCII codes of the characters in the symbols together, ignoring overflow, and taking the result modulo k or dividing it by a prime number. In fact, almost any function of the input that gives a uniform distribution of the hash values will do. Symbols can be stored by having a table consisting of k buckets numbered 0 to k − 1. All the (symbol, value) pairs whose symbol hashes to i are stored on a linked list pointed to by slot i in the hash table. With n symbols and k slots in the hash table, the average list will have length n /k.

530

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

By choosing k approximately equal to n, symbols can be located with only about one lookup on the average. By adjusting k we can reduce table size at the expense of slower lookups. Hash coding is illustrated in Fig. 7-12. Andy Anton Cathy Dick Erik Frances Frank Gerrit Hans Henri Jan Jaco Maarten Reind Roel Willem Wiebren

14025 31253 65254 54185 47357 56445 14332 32334 44546 75544 17097 64533 23267 63453 76764 34544 34344

0 4 5 0 6 3 3 4 4 2 5 6 0 1 7 6 1

(a)

Hash table

Linked table

0

Andy

14025

Maarten

23267

1

Reind

63453

Wiebren

34344

2

Henri

75544

3

Frances

56445

Frank

14332

4

Hans

44546

Gerrit

32334

5

Jan

17097

Cathy

65254

6

Jaco

64533

Willem

34544

7

Roel

76764

Dick

54185

Anton

31253

Erik

47357

(b)

Figure 7-12. Hash coding. (a) Symbols, values, and the hash codes derived from the symbols. (b) Eight-entry hash table with linked lists of symbols and values.

7.4 LINKING AND LOADING Most programs consist of more than one procedure. Compilers and assemblers generally translate one procedure at a time and put the translated output on disk. Before the program can be run, all the translated procedures must be found

SEC. 7.4

531

LINKING AND LOADING

and linked together properly. If virtual memory is not available, the linked program must be explicitly loaded into main memory as well. Programs that perform these functions are called by various names, including linker, linking loader, and linkage editor. The complete translation of a source program requires two steps, as shown in Fig. 7-13: 1. Compilation or assembly of the source procedures. 2. Linking of the object modules. The first step is performed by the compiler or assembler and the second one is performed by the linker. Source procedure 1

Source procedure 2

Source procedure 3

Object module 1

Translator

Object module 2

Linker

Executable binary program

Object module 3

Figure 7-13. Generation of an executable binary program from a collection of independently translated source procedures requires using a linker.

The translation from source procedure to object module represents a change of level because the source language and target language have different instructions and notation. The linking process, however, does not represent a change of level, since both the linker’s input and the linker’s output are programs for the same virtual machine. The linker’s function is to collect procedures translated separately and link them together to be run as a unit called an executable binary program. On MS-DOS, Windows 95/98, and NT the object modules have extension .obj and the executable binary programs have extension .exe. On UNIX , the object modules have extension .o; executable binary programs have no extension. Compilers and assemblers translate each source procedure as a separate entity for a good reason. If a compiler or assembler were to read a series of source procedures and directly produce a ready-to-run machine language program, changing one statement in one source procedure would require that all the source procedures be retranslated. If the separate-object-module technique of Fig. 7-13 is used, it is only necessary to retranslate the modified procedure and not the unchanged ones, although it is necessary to relink all the object modules again. Linking is usually much faster than translating, however; thus the two-step process of translating and linking can save a great deal of time during the development of a program. This gain is especially important for programs with hundreds or thousands of modules.

532

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

7.4.1 Tasks Performed by the Linker At the start of pass one of the assembly process, the instruction location counter is set to 0. This step is equivalent to assuming that the object module will be located at (virtual) address 0 during execution. Figure 7-14 shows four object modules for a generic machine. In this example, each module begins with a BRANCH instruction to a MOVE instruction within the module. 400

Object module A

600

300

CALL B

500

200

MOVE P TO X

400 300

100 0

500 400

BRANCH TO 200

200

Object module C

100

CALL D

300

300 200

MOVE R TO X

200

CALL C

MOVE Q TO X

BRANCH TO 300 Object module D

MOVE S TO X

100

100 0

0

Object module B

BRANCH TO 200

0

BRANCH TO 200

Figure 7-14. Each module has its own address space, starting at 0.

In order to run the program, the linker brings the object modules into main memory to form the image of the executable binary program, as shown in Fig. 715(a). The idea is to make an exact image of the executable program’s virtual address space inside the linker and position all the object modules at their correct locations. If there is not enough (virtual) memory to form the image, a disk file can be used. Typically, a small section of memory starting at address zero is used for interrupt vectors, communication with the operating system, catching uninitialized pointers, or other purposes, so programs often start above 0. In this figure we have (arbitrarily) started programs at address 100.

SEC. 7.4

LINKING AND LOADING

533

The program of Fig. 7-15(a), although loaded into the image of the executable binary file, is not yet ready for execution. Consider what would happen if execution began with the instruction at the beginning of module A. The program would not branch to the MOVE instruction as it should, because that instruction is now at 300. In fact, all memory reference instructions will fail for the same reason. Clearly something has to be done. This problem, called the relocation problem, occurs because each object module in Fig. 7-14 represents a separate address space. On a machine with a segmented address space, such as the Pentium 4, theoretically each object module could have its own address space by being placed in its own segment. However, OS/2 is the only operating system for the Pentium 4 that supports this concept. All versions of Windows and UNIX support only one linear address space, so all the object modules must be merged together into a single address space. Furthermore, the procedure call instructions in Fig. 7-15(a) will not work either. At address 400, the programmer had intended to call object module B, but because each procedure is translated by itself, the assembler has no way of knowing what address to insert into the CALL B instruction. The address of object module B is not known until linking time. This problem is called the external reference problem. Both of these problems can be solved in a simple way by the linker. The linker merges the separate address spaces of the object modules into a single linear address space in the following steps: 1. It constructs a table of all the object modules and their lengths. 2. Based on this table, it assigns a starting address to each object module. 3. It finds all the instructions that reference memory and adds to each a relocation constant equal to the starting address of its module. 4. It finds all the instructions that reference other procedures and inserts the address of these procedures in place. The object module table constructed in step 1 is shown for the modules of Fig. 7-15 below. It gives the name, length, and starting address of each module. Module A B C D

Length 400 600 500 300

Starting address 100 500 1100 1600

Figure 7-15(b) shows how the address space of Fig. 7-15(a) looks after the linker has performed these steps.

534

THE ASSEMBLY LANGUAGE LEVEL

1900 1800

1900 MOVE S TO X

1700 1600 1500

Object module D

BRANCH TO 200

MOVE R TO X

Object module C

1000

BRANCH TO 200

1100 1000

CALL C

MOVE Q TO X

Object module B

800 700

600

600 BRANCH TO 300

400

CALL B

300

MOVE P TO X

200 100 0

CALL 1600

MOVE R TO X

Object module C

BRANCH TO 1300

CALL 1100

900

700

500

BRANCH TO 1800

1400 1300

Object module D

1200

900 800

MOVE S TO X

1700

1500

1200 1100

1800

1600

CALL D

1400 1300

CHAP. 7

500

Object module A

MOVE Q TO X

BRANCH TO 800

400

CALL 500

300

MOVE P TO X

200 BRANCH TO 200

100

Object module B

BRANCH TO 300

0

Figure 7-15. (a) The object modules of Fig. 7-14 after being positioned in the binary image but before being relocated and linked. (b) The same object modules after linking and after relocation has been performed.

Object module A

SEC. 7.4

LINKING AND LOADING

535

7.4.2 Structure of an Object Module Object modules often contain six parts, as shown in Fig. 7-16. The first part contains the name of the module, certain information needed by the linker, such as the lengths of the various parts of the module, and sometimes the assembly date. End of module Relocation dictionary

Machine instructions and constants

External reference table Entry point table Identification

Figure 7-16. The internal structure of an object module produced by a translator.

The second part of the object module is a list of the symbols defined in the module that other modules may reference, together with their values. For example, if the module consists of a procedure named bigbug, the entry point table will contain the character string ‘‘bigbug’’ followed by the address to which it corresponds. The assembly language programmer indicates which symbols are to be declared as entry points by using a pseudoinstruction such as PUBLIC in Fig. 7-3. The third part of the object module consists of a list of the symbols that are used in the module but which are defined in other modules, along with a list of which machine instructions use which symbols. The linker needs the latter list in order to be able to insert the correct addresses into the instructions that use external symbols. A procedure can call other independently translated procedures by declaring the names of the called procedures to be external. The assembly language programmer indicates which symbols are to be declared as external symbols by using a pseudoinstruction such as EXTERN in Fig. 7-3. On some computers entry points and external references are combined into one table. The fourth part of the object module is the assembled code and constants. This part of the object module is the only one that will be loaded into memory to be executed. The other five parts will be used by the linker to help it do its work and then discarded before execution begins.

536

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

The fifth part of the object module is the relocation dictionary. As shown in Fig. 7-15, instructions that contain memory addresses must have a relocation constant added. Since the linker has no way of telling by inspection which of the data words in part four contain machine instructions and which contain constants, information about which addresses are to be relocated is provided in this table. The information may take the form of a bit table, with 1 bit per potentially relocatable address, or an explicit list of addresses to be relocated. The sixth part is an end-of-moduly mark, perhaps a checksum to catch errors made while reading the module, and the address at which to begin execution. Most linkers require two passes. On pass one the linker reads all the object modules and builds up a table of module names and lengths, and a global symbol table consisting of all entry points and external references. On pass two the object modules are read, relocated, and linked one module at a time.

7.4.3 Binding Time and Dynamic Relocation In a multiprogramming system, a program can be read into main memory, run for a little while, written to disk, and then read back into main memory to be run again. In a large system, with many programs, it is difficult to ensure that a program is read back into the same locations every time. Figure 7-17 shows what would happen if the already relocated program of Fig. 7-15(b) were reloaded at address 400 instead of address 100 where the linker put it originally. All the memory addresses are incorrect; moreover, the relocation information has long since been discarded. Even if the relocation information were still available, the cost of having to relocate all the addresses every time the program was swapped would be too high. The problem of moving programs that have already been linked and relocated is intimately related to the time at which the final binding of symbolic names onto absolute physical memory addresses is completed. When a program is written it contains symbolic names for memory addresses, for example, BR L. The time at which the actual main memory address corresponding to L is determined is called the binding time. At least six possibilities for the binding time exist: 1. When the program is written. 2. When the program is translated. 3. When the program is linked but before it is loaded. 4. When the program is loaded. 5. When a base register used for addressing is loaded. 6. When the instruction containing the address is executed.

SEC. 7.4

LINKING AND LOADING 2200 2100

MOVE S TO X

2000 1900 1800

BRANCH TO 1800

CALL 1600

1700 1600

Object module D

MOVE R TO X

Object module C

1500 1400

BRANCH TO 1300

1300

CALL 1100

1200 1100

MOVE Q TO X

Object module B

1000 900 800

BRANCH TO 800

700

CALL 500

600

MOVE P TO X

Object module A

500 400

BRANCH TO 300

0

Figure 7-17. The relocated binary program of Fig. 7-15(b) moved up 300 addresses. Many instructions now refer to an incorrect memory address.

537

538

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

If an instruction containing a memory address is moved after binding, it will be incorrect (assuming that the object referred to has also been moved). If the translator produces an executable binary as output, the binding has occurred at translation time, and the program must be run at the address the translator expected it to be run at. The linking method described in the preceding section binds symbolic names to absolute addresses during linking, which is why moving programs after linking fails, as shown in Fig. 7-17. Two related issues are involved here. First, there is the question of when symbolic names are bound to virtual addresses. Second, there is a question of when virtual addresses are bound to physical addresses. Only when both operations have taken place is binding complete. When the linker merges the separate address spaces of the object modules into a single linear address space, it is, in fact, creating a virtual address space. The relocation and linking serve to bind symbolic names onto specific virtual addresses. This observation is true whether or not virtual memory is being used. Assume for the moment that the address space of Fig. 7-15(b) were paged. It is clear that the virtual addresses corresponding to the symbolic names A, B, C, and D have already been determined, even though their physical main memory addresses will depend on the contents of the page table at the time they are used. An executable binary program is really a binding of symbolic names onto virtual addresses. Any mechanism that allows the mapping of virtual addresses onto physical memory addresses to be changed easily will facilitate moving programs around in main memory, even after they have been bound to a virtual address space. One such mechanism is paging. After a program has been moved in main memory, only its page table need be changed, not the program itself. A second mechanism is the use of a runtime relocation register. The CDC 6600 and its successors had such a register. On machines using this relocation technique, the register always points to the physical memory address of the start of the current program. All memory addresses have the relocation register added to them by the hardware before being sent to the memory. The entire relocation process is transparent to the user programs. They do not even know that it is occurring. When a program is moved, the operating system must update the relocation register. This mechanism is less general than paging because the entire program must be moved as a unit (unless there are separate code and data relocation registers, as on the Intel 8088, in which case it has to be moved as two units). A third mechanism is possible on machines that can refer to memory relative to the program counter. Many branch instructions are PC relative, which helps. Whenever a program is moved in main memory only the program counter need be updated. A program, all of whose memory references are either relative to the program counter or absolute (e.g., to I/O device registers at absolute addresses) is said to be position independent. A position-independent procedure can be placed anywhere within the virtual address space without the need for relocation.

SEC. 7.4

LINKING AND LOADING

539

7.4.4 Dynamic Linking The linking strategy discussed in Sec. 7.4.1 has the property that all procedures that a program might call are linked before the program can begin execution. On a computer with virtual memory, completing all linking before beginning execution does not take advantage of the full capabilities of the virtual memory. Many programs have procedures that are called only under unusual circumstances. For example, compilers have procedures for compiling rarely used statements, plus procedures for handling error conditions that seldom occur. A more flexible way to link separately compiled procedures is to link each procedure at the time it is first called. This process is known as dynamic linking. It was pioneered by MULTICS whose implementation is in some ways still unsurpassed. In the next sections we will look at dynamic linking in several systems. Dynamic Linking in MULTICS In the MULTICS form of dynamic linking, associated with each program is a segment, called the linkage segment, which contains one block of information for each procedure that might be called. This block of information starts with a word reserved for the virtual address of the procedure and it is followed by the procedure name, which is stored as a character string. When dynamic linking is being used, procedure calls in the source language are translated into instructions that indirectly address the first word of the corresponding linkage block, as shown in Fig. 7-18(a). The compiler fills this word with either an invalid address or a special bit pattern that forces a trap. When a procedure in a different segment is called, the attempt to address the invalid word indirectly causes a trap to the dynamic linker. The linker then finds the character string in the word following the invalid address and searches the user’s file directory for a compiled procedure with this name. That procedure is then assigned a virtual address, usually in its own private segment, and this virtual address overwrites the invalid address in the linkage segment, as indicated in Fig. 7-18(b). Next, the instruction causing the linkage fault is re-executed, allowing the program to continue from the place it was before the trap. All subsequent references to that procedure will be executed without causing a linkage fault, for the indirect word now contains a valid virtual address. Consequently, the dynamic linker is invoked only the first time a procedure is called and not thereafter. Dynamic Linking in Windows All versions of the Windows operating system, including NT, support dynamic linking and rely heavily on it. Dynamic linking uses a special file format called a DLL (Dynamic Link Library). DLLs can contain procedures, data, or both.

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

; ;;;; ;;;;

540

A procedure segment CALL EARTH

The linkage segment rect Indi ssing Invalid address re add E A R T H

CALL FIRE CALL AIR

CALL WATER

Indirect word

Invalid address A I R

Linkage information for the procedure of AIR

Invalid address F I R E

Name of the procedure is stored as a character string

Invalid address A A T E R

CALL EARTH

CALL WATER

(a)

A procedure segment CALL EARTH

The linkage segment rect Indi ssing Address of earth e r add E A R T H

To earth

CALL FIRE

A

CALL AIR

Invalid address I R

Invalid address

F

CALL WATER

I

R E

Invalid address W A T E R

CALL EARTH

CALL WATER

(b)

Figure 7-18. Dynamic linking. (a) Before EARTH is called. (b) After EARTH has been called and linked.

They are commonly used to allow two or more processes to share library procedures or data. Many DLLs have extension .dll, but other extensions are also in use, including .drv (for driver libraries) and .fon (for font libraries).

SEC. 7.4

541

LINKING AND LOADING

The most common form of a DLL is a library consisting of a collection of procedures that can be loaded into memory and accessed by multiple processes at the same time. Figure 7-19 illustrates two programs sharing a DLL file that contains four procedures, A, B, C, and D. Program 1 uses procedure A; program 2 uses procedure C, although they could equally well have used the same procedure. User process 1

User process 2

DLL Header A B C D

Figure 7-19. Use of a DLL file by two processes.

A DLL is constructed by the linker from a collection of input files. In fact, building a DLL file is very much like building an executable binary program, except that a special flag is given to the linker to tell it to make a DLL. DLLs are commonly constructed from collections of library procedures that are likely to be needed by multiple processes. The interface procedures to the Windows system call library and large graphics libraries are common examples of DLLs. The advantage of using DLLs is saving space in memory and on disk. If some common library were statically bound to each program using it, it would appear in many executable binaries on the disk and in memory, wasting space. With DLLs, each library only appears once on disk and once in memory. In addition to saving space, this approach makes it easy to update library procedures, even after the programs using them have been compiled and linked. For commercial software packages, where the users rarely have the source code, using DLLs means that the software vendor can fix bugs in the libraries by just distributing new DLL files over the Internet, without requiring any changes to the main program binaries. The main difference between a DLL and an executable binary is that a DLL cannot be started and run on its own (because it has no main program). It also has different information in its header. In addition, the DLL as a whole has several extra procedures not related to the procedures in the library. For example, there is

542

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

one procedure that is automatically called whenever a new process is bound to the DLL and another one that is automatically called whenever a process is unbound from it. These procedures can allocate and deallocate memory or manage other resources needed by the DLL. There are two ways for a program to bind to a DLL. In the first way, called implicit linking, the user’s program is statically linked with a special file called an import library that is generated by a utility program that extracts certain information from the DLL. The import library provides the glue that allows the user program to access the DLL. A user program can be linked with multiple import libraries. When a program using implicit linking is loaded into memory for execution, Windows examines it to see which DLLs it uses and checks to see if all of them are already in memory. Those that are not in memory are loaded immediately (but not necessarily in their entirety, since they are paged). Some changes are then made to the data structures in the import libraries so the called procedures can be located, somewhat analogous to the changes shown in Fig. 718. They also have to be mapped into the program’s virtual address space. At this point, the user program is ready to run and can call the procedures in the DLLs as though they had been statically bound with it. The alternative to implicit linking is (not surprisingly) explicit linking. This approach does not require import libraries and does not cause the DLLs to be loaded at the same time the user program is. Instead, the user program makes an explicit call at run time to bind to a DLL, then makes additional calls to get the addresses of procedures it needs. Once these have been found, it can call the procedures. When it is all done, it makes a final call to unbind from the DLL. When the last process unbinds from a DLL, the DLL can be removed from memory. It is important to realize that a procedure in a DLL does not have any identity of its own (as a thread or process does). It runs in the caller’s thread and uses the caller’s stack for its local variables. It can have process-specific static data (as well as shared data) and otherwise behaves the same as a statically-linked procedure. The only essential difference is how the binding to it is performed. Dynamic Linking in UNIX The UNIX system has a mechanism similar in essence to DLLs in Windows. It is called a shared library. Like a DLL file, a shared library is an archive file containing multiple procedures or data modules that are present in memory at run time and can be bound to multiple processes at the same time. The standard C library and much of the networking code are shared libraries. UNIX supports only implicit linking, so a shared library consist of two parts: a host library, which is statically linked with the executable file, and a target library, which is called at run time. While the details differ, the concepts are essentially the same as with DLLs.

SEC. 7.5

SUMMARY

543

7.5 SUMMARY Although most programs can and should be written in a high-level language, occasional situations exist in which assembly language is needed, at least in part. Programs for resource-poor portable computers such as smart cards, embedded processors in appliances, and wireless portable digital assistants are potential candidates. An assembly language program is a symbolic representation for some underlying machine language program. It is translated to the machine language by a program called an assembler. When extremely fast execution is critical to the success of some application, a better approach than writing everything in assembly language is to first write the whole program in a high-level language, then measure where it is spending its time, and finally rewrite only those portions of the program that are heavily used. In practice, a small fraction of the code is usually responsible for a large fraction of the execution time. Many assemblers have a macro facility that allows the programmer to give commonly used code sequences symbolic names for subsequent inclusion. Usually, these macros can be parameterized in a straightforward way. Macros are implemented by a kind of literal string-processing algorithm. Most assemblers are two pass. Pass one is devoted to building up a symbol table for labels, literals, and explicitly declared identifiers. The symbols can either be kept unsorted and then searched linearly, first sorted and then searched using binary search, or hashed. If symbols do not need to be deleted during pass one, hashing is usually the best method. Pass two does the code generation. Some pseudoinstructions are carried out on pass one and some on pass two. Independently-assembled programs can be linked together to form an executable binary program that can be run. This work is done by the linker. Its primary tasks are relocation and binding of names. Dynamic linking is a technique in which certain procedures are not linked until they are actually called. Windows DLLs and UNIX shared libraries use dynamic linking.

PROBLEMS 1. For a certain program, 2% of the code accounts for 50% of the execution time. Compare the following three strategies with respect to programming time and execution time. Assume that it would take 100 man-months to write it in C, and that assembly code is 10 times slower to write and four times more efficient. a. Entire program in C. b. Entire program in assembler. c. First all in C, then the key 2% rewritten in assembler.

544

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

2. Do the considerations that hold for two-pass assemblers also hold for compilers? a. Assume that the compilers produce object modules, not assembly code. b. Assume that the compilers produce symbolic assembly language. 3. Most assemblers for the Intel CPUs have the destination address as the first operand and the source address as the second operand. What problems would have to be solved to do it the other way? 4. Can the following program be assembled in two passes? EQU is a pseudoinstruction that equates the label to the expression in the operand field. P EQU Q Q EQU R R EQU S S EQU 4

5. The Dirtcheap Software Company is planning to produce an assembler for a computer with a 48-bit word. To keep costs down, the project manager, Dr. Scrooge, has decided to limit the length of allowed symbols so that each symbol can be stored in a single word. Scrooge has declared that symbols may consist only of letters, except the letter Q, which is forbidden (to demonstrate their concern for efficiency to the customers). What is the maximum length of a symbol? Describe your encoding scheme. 6. What is the difference between an instruction and a pseudoinstruction? 7. What is the difference between the instruction location counter and the program counter, if any? After all, both keep track of the next instruction in a program. 8. Show the symbol table after the following Pentium 4 statements have been encountered. The first statement is assigned to address 1000. EVEREST: K2: WHITNEY: MCKINLEY: FUJI: KIBO:

POP BX PUSH BP MOV BP,SP PUSH X PUSH SI SUB SI,300

(1 BYTE) (1 BYTE) (2 BYTES) (3 BYTES) (1 BYTE) (3 BYTES)

9. Can you envision circumstances in which an assembly language permits a label to be the same as an opcode (e.g., MOV as a label)? Discuss. 10. Show the steps needed to look up Berkeley using binary search on the following list: Ann Arbor, Berkeley, Cambridge, Eugene, Madison, New Haven, Palo Alto, Pasadena, Santa Cruz, Stony Brook, Westwood, and Yellow Springs. When computing the middle element of a list with an even number of elements, use the element just after the middle index. 11. Is it possible to use binary search on a table whose size is prime?

CHAP. 7

PROBLEMS

545

12. Compute the hash code for each of the following symbols by adding up the letters (A = 1, B = 2, etc.) and taking the result modulo the hash table size. The hash table has 19 slots, numbered 0 to 18. els, jan, jelle, maaike Does each symbol generate a unique hash value? If not, how can the collision problem be dealt with? 13. The hash coding method described in the text links all the entries having the same hash code together on a linked list. An alternative method is to have only a single nslot table, with each table slot having room for one key and its value (or pointers to them). If the hashing algorithm generates a slot that is already full, a second hashing algorithm is used to try again. If that one is also full, another is used, and so on, until an empty is found. If the fraction of the slots that are full is R, how many probes will be needed, on the average, to enter a new symbol? 14. As technology progresses, it may one day be possible to put thousands of identical CPUs on a chip, each CPU having a few words of local memory. If all CPUs can read and write three shared registers, how can an associative memory be implemented? 15. The Pentium 4 has a segmented architecture, with multiple independent segments. An assembler for this machine might well have a pseudoinstruction SEG N that would direct the assembler to place subsequent code and data in segment N. Does this scheme have any influence on the ILC? 16. Programs often link to multiple DLLs. Would it not be more efficient just to put all the procedures in one big DLL and then link to it? 17. Can a DLL be mapped into two process’ virtual address spaces at different virtual addresses? If so, what problems arise? Can they be solved? If not, what can be done to eliminate them? 18. One way to do (static) linking is as follows. Before scanning the library, the linker builds a list of procedures needed, that is, names defined as EXTERN in the modules being linked. Then the linker goes through the library linearly, extracting every procedure that is in the list of names needed. Does this scheme work? If not, why not and how can it be remedied? 19. Can a register be used as the actual parameter in a macro call? How about a constant? Why or why not? 20. You are to implement a macro assembler. For esthetic reasons, your boss has decided that macro definitions need not precede their calls. What implications does this decision have on the implementation? 21. Think of a way to put a macro assembler into an infinite loop. 22. A linker reads five modules, whose lengths are 200, 800, 600, 500, and 700 words, respectively. If they are loaded in that order, what are the relocation constants? 23. Write a symbol table package consisting of two routines: enter(symbol, value) and lookup(symbol, value). The former enters new symbols in the table and the latter looks them up. Use some form of hash coding.

546

THE ASSEMBLY LANGUAGE LEVEL

CHAP. 7

24. Write a simple assembler for the Mic-1 computer of Chap. 4. In addition to handling the machine instructions, provide a facility for assigning constants to symbols at assembly time, and a way to assemble a constant into a machine word. 25. Add a simple macro facility to the assembler of the preceding problem.

8 PARALLEL COMPUTER ARCHITECTURES

Although computers keep getting faster, the demands placed on them are increasing at least as fast. Astronomers want to simulate the entire history of the universe, from the big bang until the show is over. Pharmaceutical engineers would love to be able to design medicines to order for specific diseases on their computers instead of having to sacrifice legions of rats. Aircraft designers could come up with more fuel-efficient products if computers could do all the work, without the need for constructing physical wind tunnel prototypes. In short, however much computing power is available, for many users, especially in science, engineering, and industry, it is never enough. Although clock rates are continually rising, circuit speed cannot be increased indefinitely. The speed of light is already a major problem for designers of highend computers, and the prospects of getting electrons and photons to move faster are dim. Heat-dissipation issues are turning supercomputers into state-of-the-art air conditioners. Finally, as transistor sizes continue to shrink, at some point each transistor will have so few atoms in it that quantum mechanical effects (e.g., the Heisenberg uncertainty principle) may become a major problem. Therefore, in order to handle larger and larger problems, computer architects are turning increasingly to parallel computers. While it may not be possible to build a computer with one CPU and a cycle time of 0.001 nsec, it may well be possible to build one with 1000 CPUs each with a cycle time of 1 nsec. Although the latter design uses slower CPUs than the former one, its total computing capacity is theoretically the same. Herein lies the hope. 547

548

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Parallelism can be introduced at various levels. At the lowest level, it can be added to the CPU chip, by pipelining and superscalar designs with multiple functional units. It can also be added by having very long instruction words with implicit parallelism. Special features can be added to a CPU to allow it to handle multiple threads of control at once. Finally, multiple CPUs can be put together on the same chip. Together, these features can pick up perhaps a factor of 10 in performance over purely sequential designs. At the next level, extra CPU boards with additional processing capacity can be added to a system. Usually, these plug-in CPUs have specialized functions, such as network packet processing, multimedia processing, or cryptography. For specialized applications, they can also gain a factor of perhaps 5 to 10. However, to win a factor of a hundred or a thousand or a million, it is necessary to replicate entire CPUs and to make them all work together efficiently. This idea leads to large multiprocessors and multicomputers (cluster computers). Needless to say, hooking up thousands of processors into a big system leads to its own problems that need to be solved. Finally, it is now possible to lash together entire organizations over the Internet to form very loosely coupled compute grids. These systems are only starting to emerge, but have interesting potential for the future. When two CPUs or processing elements are close together, have a high bandwidth and low delay between them, and are computationally intimate, they are said to be tightly coupled. Conversely, when they are far apart, have a low bandwidth and high delay and are computationally remote, they are said to be loosely coupled. In this chapter we will look at the design principles for these various forms of parallelism and study a variety of examples. We will start with the most tightly-coupled systems, those that use on-chip parallelism, and gradually move to more and more loosely coupled systems, ending with a few words on grid computing. This spectrum is crudely illustrated in Fig. 8-1. The whole issue of parallelism, from one end of the spectrum to the other, is a hot topic of research. Accordingly, many more references are given in this chapter, primary to recent papers on the subject. More introductory references are given in Sec. 9.1.8.

8.1 ON-CHIP PARALELLISM One way to increase the throughput of a chip is to have it do more things at the same time. In this section, we will look at some of the ways of achieving speedup through parallelism at the chip level, including instruction-level parallelism, multithreading, and putting more than one CPU on the chip. These techniques are quite different, but each of them helps in its own way. But in all cases the idea is to get more activity going at the same time.

SEC. 8.1

549

ON-CHIP PARALELLISM Private memory Shared memory Coprocessor

Thread

CPU

CPU

Computer

M CPU Internet

M M CPU

CPU

CPU

Main CPU Tightly coupled (a)

Loosely coupled (b)

(c)

(d)

(e)

Figure 8-1. (a) On-chip parallelism. (b) A coprocessor. (c) A multiprocessor. (d) A multicomputer. (e) A grid.

8.1.1 Instruction-Level Parallelism At the lowest level, one way to achieve parallelism is to issue multiple instructions per clock cycle. Multiple-issue CPUs come in two varieties: superscalar processors and VLIW processors. We have actually touched on both earlier in the book, but it may be useful to review this material here. We have seen superscalar CPUs before (e.g., Fig. 2-5). In the most common configuration, at a certain point in the pipeline an instruction is ready to be executed. Superscalar CPUs are capable of issuing multiple instructions to the execution units in a single clock cycle. The number of instructions actually issued depends on both the processor design and on current circumstances. The hardware determines the maximum number that can be issued, usually two to six instructions. However, if an instruction needs a functional unit that is not available or a result that has not yet been computed, the instruction will not be issued. The other form of instruction-level parallelism is found in VLIW (Very Long Instruction Word) processors. In the original form, VLIW machines indeed had long words containing instructions that used multiple functional units. Consider, for example, the pipeline of Fig. 8-2(a), where the machine has five functional units and can perform two integer operations, one floating-point operation, one load, and one store simultaneously. A VLIW instruction for this machine would contain five opcodes and five pairs of operands, one opcode and operand pair per functional unit. With 6 bits per opcode, 5 bits per register, and 32 bits per memory address, instructions could easily be 134 bits—quite long indeed.

550

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Integer

Integer (a)

Fetch

Decode

Floating point

Issue

Retire

Load

Store No-op (b)

---L-

---L-

l--L-

llF-S

l-FLS

L L lL

llFS

lIF-S

---L-

End-of-bundle marker

VLIW instruction (c)

l-FL-

lFLS

lFL

lIFS

L

LFS

Bundle Figure 8-2. (a) A CPU pipeline. (b) A sequence of VLIW instructions. (c) An instruction stream with bundles marked.

However, this design proved too rigid because not every instruction was able to utilize every functional unit, leading to many useless NO-OPs used as filler, as illustrated in Fig. 8-2(b). Consequently, modern VLIW machines have a way of marking a bundle of instructions as belonging together, for example with an ‘‘end of bundle’’ bit, as shown in Fig. 8-2(c). The processor can then fetch the entire bundle and issue it all at once. It is up to the compiler to prepare bundles of compatible instructions. In effect, VLIW shifts the burden of determining which instructions can be issued together from run time to compile time. Not only does this choice make the hardware simpler and faster, but since an optimizing compiler can run for a long time if need be, better bundles can be assembled than what the hardware could do at run time. Of course, such a radical change in CPU architecture will be difficult to introduce, as demonstrated by the slow acceptance of the Itanium. It is worth noting in passing that instruction-level parallelism is not the only form of low-level parallelism. Another is memory-level parallelism, in which multiple memory operations are in flight at the same time (Chou et al., 2004).

--LFS

SEC. 8.1

551

ON-CHIP PARALELLISM

The TriMedia VLIW CPU We studied one example of a VLIW CPU, the Itanium-2, in Chap. 5. Let us now look at a very different VLIW processor, the TriMedia, designed by Philips, the Dutch electronics company that also invented the audio CD and CD-ROM. The TriMedia is intended for use as an embedded processor in image-, audio-, and video-intensive applications such as CD, DVD, and MP3 players, CD and DVD recorders, interactive TV sets, digital cameras, camcorders, and so on. Given these application areas, it is not surprising that it differs considerably from the Itanium-2, which is a general-purpose CPU intended for high-end servers. The TriMedia is a true VLIW processor with every instruction holding as many as five operations. Under completely optimal conditions, every clock cycle one instruction is started and the five operations are issued. The clock runs at 266 MHz or 300 MHz, but since it can issue five operations per cycle, the effective clock speed is as much as five times higher. In the discussion below, we will focus on the TM3260 implementation of the TriMedia; other versions differ in minor ways from it. A typical instruction is illustrated in Fig. 8-3. The instructions vary from standard 8-, 16-, and 32-bit integer instructions, through IEEE 754 floating-point instructions to parallel multimedia instructions. As a consequence of the five issues per cycle and the parallel multimedia instructions, the TriMedia is fast enough to decode streaming DV from a camcorder at full size and full frame rate in software. Slot 1 Addition

Slot 2 Shift

Slot 3 Multimedia

Slot 4 Load

Operation in slot 5 Store

Instruction Figure 8-3. A typical TriMedia instruction, showing five possible operations.

The TriMedia has a byte-oriented memory, with the I/O registers mapped into the memory space. Half words (16 bits) and full words (32 bits) must be aligned on their natural boundaries. It can run either as little endian or big endian, depending on a PSW bit that the operating system can set. This bit affects only the way load operations and store operations transfer between memory and registers. The CPU contains a split 8-way set-associative cache, with a 64-byte line size for both the instruction cache and the data cache. The instruction cache is 64 KB; the data cache is 16 KB. There are 128 general-purpose 32-bit registers. Register R0 is hardwired to 0. Register R1 is hardwired to 1. Attempting to change either one gives the CPU a heart attack. The remaining 126 registers are all functionally equivalent and can

552

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

be used for any purpose. In addition, four special-purpose, 32-bit registers also exist. These are the program counter, program status word, and two registers that relate to interrupts. Finally, a 64-bit register counts the number of CPU cycles since the CPU was last reset. At 300 MHz, it takes nearly 2000 years for the counter to wrap around. The Trimedia TM3260 has 11 different functional units for doing arithmetic, logical, and control flow operations (as well as one for cache control that we will not discuss here). They are listed in Fig. 8-4. The first two columns name the unit and give a brief description of what it does. The third column tells how many hardware copies of the unit exist. The fourth column gives the latency, that is, how many clock cycles it takes to complete. In this context, it is worth nothing that all the functional units except the FP square root/divide unit are pipelined. The latency given in the table tells how long before the result of an operation is available, but a new operation can be initiated every cycle. Thus, for example, each of three consecutive instructions can hold two load operations, resulting in six loads in various stages of execution at the same time. Finally, the last five columns show which instruction slots can be used for which functional unit. For example, floating-point compare operations must appear only in the third slot of an instruction # ################################################################################### ! ! # ! Lat. ! 1 ! 2 ! 3 ! 4 ! 5 ! ! Unit Description #! ################################################################################### ! ! ! ! ! ! ! ! ! ! Immediate operations ! 5 ! 1 ! x ! x ! x ! x ! x ! !#Constant ################################################################################### ! 5 ! 1 ! x ! x ! x ! x ! x ! ! ALU ! 32-Bit arithmetic, Boolean ops #! Integer ################################################################################### ! ! ! ! ! ! ! ! ! ! Multibit shifts ! 2 ! 1 ! x ! x ! x ! x ! x ! #! Shifter ################################################################################### ! ! ! ! ! ! ! ! ! ! ################################################################################### ! Memory operations ! 2 ! 3 ! ! ! ! x ! x ! !#Load/Store ! 2 ! 3 ! ! x ! x ! ! ! !#Int/FP MUL ! 32-Bit integer and FP multiplies ################################################################################### ! ! ! ! ! ! ! ! ! ! ALU ! FP arithmetic ! 2 ! 3 ! x ! ! ! x ! ! !#FP ################################################################################### ! 1 ! 1 ! ! ! x ! ! ! ! FP compare ! FP compares # ################################################################################### ! ! ! ! ! ! ! ! ! ! sqrt/div ! FP division and square root ! 1 ! 17 ! ! x ! ! ! ! !#FP ################################################################################### ! Control flow ! 3 ! 3 ! ! x ! x ! x ! ! ! Branch ! ! ! ! ! ! ! ! ! !# ################################################################################### ! Dual 16-bit, quad 8-bit multimedia arithmetic ! 2 ! 3 ! x ! ! x ! ! x ! !#DSP ALU ################################################################################### ! ! ! ! ! ! ! ! ! ! MUL !#DSP ################################################################################### ! Dual 16-bit, quad 8-bit multimedia multiplies ! 2 ! 3 ! ! x ! x ! ! ! Figure 8-4. The TM3260 functional units, their quantity, latency, and which instruction slots they can use.

The constant unit is used for immediate operations, such as loading a number stored in the operation itself into a register. The integer ALU does addition, subtraction, the usual Boolean operations, and pack/unpack operations. The shifter can shift a register in either direction a specified number of bits. The load/store unit fetches memory words into registers and writes them back. The TriMedia is basically an augmented RISC CPU, so normal operations operate

SEC. 8.1

ON-CHIP PARALELLISM

553

on registers and the load/store unit is used to access memory. Transfers can be 8, 16, or 32 bits. Arithmetic and logical instructions do not access memory. The multiply unit handles both integer and floating-point multiplications. The next three units handle floating-point additions/subtractions, compares, and square roots and divisions, respectively. Branch operations are executed by the branch unit. There is a fixed 3-cycle delay after a branch, so the three instructions (up to 15 operations) following a branch are always executed, even for unconditional branches. Finally, we come to the two multimedia units, which handle the special multimedia operations. The DSP in the name of the functional unit refers to Digital Signal Processor, which the multimedia operations effectively replace. We will describe the multimedia operations briefly below. One noteworthy feature is that they all use saturated arithmetic instead of twos complement arithmetic used by the integer operations When an operation produces a result that cannot be expressed due to overflow, instead of generating an exception or giving a garbage result, the closest valid number is used. For example, with 8-bit unsigned numbers, adding 130 and 130 gives 255. Because not every operation can appear in every slot, it frequently happens that an instruction does not contain all five potential operations. When a slot is not used, it is compacted to minimized the amount of space wasted. Operations that are present occupy 26, 34, or 42 bits. Depending on the number of operations actually present, TriMedia instructions vary from 2 to 28 bytes, including some fixed overhead. The TriMedia does not make runtime checks to see if the operations in an instruction are compatible. If they are not, it just executes them anyway and gets the wrong answer. Leaving the check out was a deliberate decision to save a time and transistors. The Pentium does do runtime checking to make sure all the superscalar operations are compatible, but at a huge cost in complexity, time, and transistors. The TriMedia avoids this expense by putting the burden of scheduling on the compiler, which has all the time in the world to carefully optimize the placement of operations in instruction words. On the other hand, if an operation needs a functional unit that is not available, the instruction will stall until it becomes available. As in the Itanium-2, TriMedia operations are predicated. Each operation (with two minor exceptions) specifies a register that is tested before the operation is executed. If the low-order bit of the register is set, the operation is executed; otherwise, it is skipped. Each of the (up to) five operations is individually predicated. An example of a predicated operation is IF R2 IADD R4, R5 –> R8

which tests R2 and, if the low-order bit is 1, adds R4 to R5 and stores the result in R8. An operation can be made unconditional by using R1 (which is always 1) as the predicate register. Using R0 (which is always 0) makes it a no-op.

554

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

The TriMedia multimedia operations can be grouped into the 15 groups listed in Fig. 8-5. Many of the operations involving clipping, which specifies an operand and a range, and forces the operand into the range, using the lowest or highest values for operands that fall outside it. Clipping can be done on 8-, 16-, or 32-bit operands. For example, when clipping is performed with a range of 0 to 255 on 40 and 340, the clipped results are 40 and 255, respectively. The clip group performs clip operations. ##################################################################### ! ! ! Group Description #! #################################################################### ! ! ! Clip 4 bytes or 2 halfwords ! #! Clip #################################################################### ! DSP absolute value ! Clipped, signed, absoluted value ! #! #################################################################### ! ! ! Clipped signed addition ! add #! DSP #################################################################### ! ! ! DSP subtract Clipped signed subtraction !##################################################################### ! ! !#DSP ! Clipped signed multiplication ! multiply #################################################################### ! ! ! max !#Min, ! Get minimum or maximum of four byte pairs ! #################################################################### ! Compare ! Bytewise compare of two registers ! !##################################################################### ! ! !#Shift ! Shift a pair of 16-bit operands ! #################################################################### ! Sum of products ! Signed sum of 8- or 16-bit products ! !##################################################################### ! ! !#Merge, ! pack, swap ! Byte and halfword manipulation #################################################################### ! ! ! quad averages ! Unsigned byte-wise quad averaging #################################################################### !#Byte ! ! Byte averages ! Unsigned byte-wise average of four elements ! # #################################################################### ! ! ! multiplies !#Byte ! Unsigned 8-bit multiply ! #################################################################### ! estimation ! Unsigned sum of absolute values of signed 8-bit diffs ! #! Motion #################################################################### ! ! !!#Miscellaneous !! Other arithmetic operations !! #################################################################### Figure 8-5. The major groups of TriMedia custom operations.

The next four groups in Fig. 8-5 perform the indicated operation on operands of various sizes, clipping the results into a specific range. The min, max group examines two registers and for each byte finds the smallest or largest value. Similarly, the compare group regards two registers as four pairs of bytes and compares each pair. Multimedia operations are rarely performed on 32-bit integers because most images are composed of RGB pixels with 8-bit values for each of the red, green, and blue colors. When an image is being processed (e.g., compressed), it is normally represented by three components, one for each color (RGB space) or a logically equivalent form (YUV space, discussed later in this chapter). Either way, a lot of processing is done on rectangular arrays containing 8-bit unsigned integers. The TriMedia has a large number of operations specifically designed for processing arrays of 8-bit unsigned integers efficiently. As a simple example, consider the upper left-hand corner of an array of 8-bit values stored in (big endian)

SEC. 8.1

555

ON-CHIP PARALELLISM

memory as illustrated in Fig. 8-6(a). The 4 × 4 block shown in the corner contains 16 8-bit values labeled A through P. Suppose, for example, that the image needs to be transposed, to produce Fig. 8-6(b). How can this task be achieved? 32-Bit register A E I M

B F J N

C G K O

(a)

D H L P

A B C D

E F G H

I J K L

(b)

M N O P

R2 A B C D

R2 A E I M

R3 E F G H

R3 B F J N

R4

J K L

R4 C G K O

R5 M N O P

R5 D H L P

I

(c)

(d)

Figure 8-6. (a) An array of 8-bit elements. (b) The transposed array. (c) The original array fetched into four registers. (d) The transposed array in four registers.

One way to do the transposition is to use 12 operations that each load a byte into a different register, followed by 12 more operations that each store a byte in its correct locations. (Note: the four bytes on the diagonal do not move in the transposition.) The problem with this approach is that it requires 24 (long and slow) operations that reference memory. An alternative approach is to start with four operations that each load a word into four different registers, R2 through R5, as shown in Fig. 8-6(c). Then the four output words are assembled by masking and shifting operations to achieve the desired output, as shown in Fig. 8-6(d). Finally, these words are stored in memory. Although this way of doing it reduces the number of memory references from 24 to 8, the masking and shifting is expensive due to the many operations required to extract and insert each byte in the correct position. The TriMedia provides a better solution than either of these. It begins by fetching the four words into registers. However, instead of building the output using masking and shifting, special operations that extract and insert bytes within registers are used to build the output. The result is that with eight memory references and eight of these special multimedia operations, the transposition can be accomplished. The code first contains an operation with two load operations in slots 4 and 5, respectively, to load words into R2 and R3, followed by another such operation to load R4 and R5. The instructions holding these operations can use slots 1, 2, and 3 for other purposes. When all the words have been loaded, the eight special multimedia operations can be packed into two instructions to build the output, followed by two operations to store them. All in all, only six instructions are needed, and 14 of the 30 slots in these instructions are available for other operations. In effect, the entire job can be done with the effective equivalent of about three or so instructions. Other multimedia operations are equally efficient.

556

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Between these powerful operations and the five issue slots per instruction, the TriMedia is highly efficient at doing the kinds of calculations needed in multimedia processing.

8.1.2 On-Chip Multithreading All modern, pipelined CPUs have an inherent problem: when a memory reference misses the level 1 and level 2 caches, there is a long wait until the requested word (and its associated cache line) are loaded into the cache, so the pipeline stalls. One approach to dealing with this situation, called on-chip multithreading, allows the CPU to manage multiple threads of control at the same time in an attempt to mask these stalls. In short, if thread 1 is blocked, the CPU still has a chance of running thread 2 in order to keep the hardware fully occupied. Although the basic idea is fairly simple, multiple variants exist, which we will now examine. The first approach, called fine-grained multithreading, is illustrated in Fig. 8-7 for a CPU with the ability to issue one instruction per clock cycle. In Fig. 8-7(a)–(c), we see three threads, A, B, and C, for 12 machine cycles. During the first cycle, thread A executes instruction A1. This instruction completes in one cycle, so in the second cycle instruction A2 is started. Unfortunately, this instruction misses on the level 1 cache so two cycles are wasted while the word needed is fetched from the level 2 cache. The thread continues in cycle 5. Similarly, threads B and C also stall occasional as well, as illustrated in the figure. In this model if an instruction stalls, subsequent instructions cannot be issued. Of course, with a more sophisticated scoreboard, sometimes new instructions can still be issued, but we will ignore that possibility in this discussion. (a) A1 A2 (b) B1

A3 A4 A5 B2

(c) C1 C2 C3 C4 Cycle

A6 A7 A8

(d) A1 B1 C1 A2 B2 C2 A3 B3 C3 A4 B4 C4

B3 B4 B5 B6 B7 B8 C5 C6

C7 C8

(e) A1 A2

B1

C1 C2 C3 C4 A3 A4 A5

Cycle

Figure 8-7. (a)–(c) Three threads. The empty boxes indicated that the thread has stalled waiting for memory. (d) Fine-grained multithreading. (e) Coarsegrained multithreading.

Fine-grained multithreading masks the stalls by running the threads round robin, with a different thread in consecutive cycles, as shown in Fig. 8-7(d). By the time the fourth cycle comes up, the memory operation initiated in A1 has completed, so instruction A2 can be run, even if it needs the result of A1. In this case the maximum stall is two cycles, so with three threads the stalled operation always completes in time. If a memory stall took four cycles, we would need four threads to insure continuous operation, and so on.

SEC. 8.1

ON-CHIP PARALELLISM

557

Since different threads have nothing to do with one another, each one needs its own set of registers. When an instruction is issued, a pointer to its register set has to be included along with the instruction so that if a register is referenced, the hardware will know which register set to use. Consequently, the maximum number of threads that can be run at once is fixed at the time the chip is designed. Memory operations are not the only reason for stalling. Sometimes an instruction needs a result computed by a previous instruction that is not yet complete. Sometimes an instruction cannot start because it follows a conditional branch whose direction is not yet known. In general, if the pipeline has k stages but there are at least k threads to run round robin, there will never be more than one instruction per thread in the pipeline at any moment so no conflicts can occur. In this situation, the CPU can run at full speed, never stalling. Of course, there may not be as many threads available as there are pipeline stages, so some designers prefer a different approach, known as coarse-grained multithreading, and illustrated in Fig. 8-7(e). Here thread A starts and continues to issue instructions until it stalls, wasting one cycle. At that point a switch occurs and B1 is executed. Since the first instruction of thread B stalls, another thread switch happens and C1 is executed in cycle 6. Since a cycle is lost whenever an instruction stalls, coarse-grained multithreading is potentially less efficient than fine-grained multithreading, but it has the big advantage that many fewer threads are needed to keep the CPU busy. In situations with an insufficient number of active threads, to be sure of finding a runnable one, coarse-grained multithreading works better. Although we have depicted coarse-grained multithreading as doing thread switches on a stall, that is not the only option. Another possibility is to switch immediately on any instruction that might cause a stall, such as a load, store, or branch, before even finding out if it does cause a stall. The latter strategy allows a switch to occur earlier (as soon as the instruction is decoded), and may make it possible to avoid dead cycles. In effect, it is saying: ‘‘Run until there might be a problem, then switch just in case.’’ Doing so makes coarse-grained multithreading somewhat more like fine-grained multithreading with its frequent switches. No matter which kind of multithreading is used, some way is needed to keep track of which operation belongs to which thread. With fine-grained multithreading, the only serious possibility is to attach a thread identifier to each operation, so as it moves through the pipeline, its identity is clear. With coarse-grained multithreading, another possibility exists: when switching threads, let the pipeline clear and only then start the next thread. In that way, only one thread at a time is in the pipeline and its identity is never in doubt. Of course, letting the pipeline run dry on a thread switch only makes sense if thread switches take place at intervals very much longer than the time it takes to empty the pipeline. So far we have assumed that the CPU can issue only one instruction per cycle. As we have seen, however, modern CPUs can issue multiple instructions per cycle. In Fig. 8-8 we assume the CPU can issue two instructions per clock cycle,

558

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

but we maintain the rule that when an instruction stalls, subsequent instructions cannot be issued. In Fig. 8-8(a) we see how fine-grained multithreading works with a dual-issue superscalar CPU. For thread A, the first two instructions can be issued in the first cycle, but for thread B we immediately hit a problem in the next cycle, so only one instruction can be issued, and so on. A1 B1 C1 A3 B2 C3 A5 B3 C5 A6 B5 C7

A1 B1 C1 C3 A3 A5 B2 C5 A6 A8 B3 B5

A2

A2

C2 A4

C4

B4 C6 A7 B6 C8

C2 C4 A4

C6 A7

Cycle

Cycle

(a)

(b)

B4 B6

A1 B1 C2 C4 A4 B2 C6 A7 B3 B5 B7 C7 A2 C1 C3 A3 A5 C5 A6 A8 B4 B6 B8 C8 Cycle (c)

Figure 8-8. Multithreading with a dual-issue superscalar CPU. (a) Fine-grained multithreading. (b) Coarse-grained multithreading. (c) Simultaneous multithreading.

In Fig. 8-8(b), we see how coarse-grained multithreading works with a dualissue CPU, but now with a static scheduler that does not introduce a dead cycle after an instruction that stalls. Basically, the threads are run in turn, with the CPU issuing two instructions per thread until it hits one that stalls, at which point it switches to the next thread at the start of the next cycle. With superscalar CPUs, a third possible way of doing multithreading is available, called simultaneous multithreading and illustrated in Fig. 8-8(c). This approach can be seen as a refinement to coarse-grained multithreading, in which a single thread is allowed to issue two instructions per cycle as long as it can, but when it stalls, instructions are immediately taken from the next thread in sequence to keep the CPU fully occupied. Simultaneous multithreading can also help keep all the functional units busy. When an instruction cannot be started because a functional unit it needs is occupied, an instruction from a different thread can be chosen instead. In this figure, we are assuming that B8 stalls in cycle 11, so C7 is started in cycle 12. For more information about multithreading, see (Dean, 2004; Kalla et al., 2004; and Kapil et al., 2004). The combination of multithreading and speculative execution is examined in Sohi and Roth (2001). Hyperthreading on the Pentium 4 Having looked at multithreading in the abstract, let us now consider a practical example of multithreading: the Pentium 4 CPU. After the Pentium 4 was already in production, the architects at Intel looked for various ways to speed it up

SEC. 8.1

ON-CHIP PARALELLISM

559

without changing the programmers’ interface, something that would never have been accepted. Five ways quickly popped up: 1. Increasing the clock speed. 2. Putting two CPUs on a chip. 3. Adding functional units. 4. Making the pipeline longer. 5. Using multithreading. An obvious way to improve performance is to increase the clock speed without changing anything else. Doing this is relatively straightforward and well understood, so each new chip that comes out is generally slightly faster than its predecessor. Unfortunately, a faster clock also has two main drawbacks that limit how much of an increase can be tolerated. First, a faster clock uses more energy, which is a huge problem for notebook computers and other battery-powered devices. Second, the extra energy input means the chip gets hotter and there is more heat to dissipate. Putting two CPUs on a chip is relatively straightforward, but it comes close to doubling the chip area if each one has its own caches and thus reduces the number of chips per wafer by a factor of two, which essentially doubles the unit manufacturing cost. If the two chips share a common cache as big as the original one, the chip area is not doubled, but cache size per CPU is halved, cutting into performance. Also, while high-end server applications can often fully utilize multiple CPUs, not all desktop applications have enough inherent parallelism to warrant two full CPUs. Adding additional functional units is also fairly easy, but it is important to get the balance right. Having 10 ALUs does little good if the chip is incapable of feeding instructions into the pipeline fast enough to keep them all busy. A longer pipeline with more stages, each doing a smaller piece of work in a shorter time period, also increases performance, but also increases the negative effects of branch mispredictions, cache misses, interrupts, and other factors that interrupt normal pipeline flow. Furthermore, to take full advantage of a longer pipeline, the clock speed has to be increased, which means more energy is consumed and more heat is produced. Finally, multithreading can be added. Its value is in having a second thread utilize hardware that would otherwise have lain fallow. After some experimentation, it became clear that a 5% increase in chip area for multithreading support gave a 25% performance gain in many applications, making this a good choice. Intel’s first multithreaded CPU was the Xeon in 2002, but multithreading was later added to the Pentium 4, starting with the 3.06 GHz version and continuing with faster versions of the Pentium processor. Intel calls the implementation of multithreading used in the Pentium 4 hyperthreading.

560

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

The basic idea is to allow two threads (or possibly processes, since the CPU cannot tell what is a thread and what is a process) to run at once. To the operating system, a hyperthreaded Pentium 4 chip looks like a dual processor in which both CPUs share a common cache and main memory. The operating system schedules the threads independently. If two applications are running at the same time, the operating system can run each one at the same time. For example, if a mail daemon is sending or receiving e-mail in the background while a user is interacting with some program in the foreground, the daemon and the user program can be run in parallel, as though there were two CPUs available. Application software that has been designed to run as multiple threads can use both virtual CPUs. For example, video editing programs usually allow users to specify certain filters to apply to each frame in some range. These filters can modify the brightness, contrast, color balance, or other properties of each frame. The program can then assign one CPU to process the even-numbered frames and the other CPU to process the odd-numbered frames, and the two of them can run completely independently of each other. Since the two threads share all the hardware resources, a strategy is needed to manage the sharing. Intel identified four useful strategies for resource sharing in conjunction with hyperthreading: resource duplication, partitioned resources, threshold sharing, and full sharing. We will now touch on each of these in turn. To start with, some resources are duplicated just for threading. For example, since each thread has its own flow of control, a second program counter had to be added. In addition, the table that maps the architectural registers (EAX, EBX, etc.) onto the physical registers also had to be duplicated, as did the interrupt controller, since the threads can be independently interrupted. Next we have partitioned resource sharing, in which the hardware resources are rigidly divided between the threads. For example, if the CPU has a queue between two functional pipeline stages, half the slots could be dedicated to thread 1 and the other half to thread 2. Partitioning is easy to accomplish, has no overhead, and keeps the threads out of each other’s hair. If all the resources are partitioned, we effectively have two separate CPUs. On the down side, it can easily happen that at some point one thread is not using some of its resources that the other one wants but is forbidden from accessing. As a consequence, resources that could have been used productively lie idle. The opposite of partitioned sharing is full resource sharing. When this scheme is used, either thread can acquire any resources it needs, first come, first served. However, imagine a fast thread consisting primarily of additions and subtractions and a slow thread consisting primarily of multiplications and divisions. If instructions are fetched from memory faster than multiplications and divisions can be carried out, the backlog of instructions fetched for the slow thread and queued but not yet fed into the pipeline will grow in time. Eventually, this backlog will occupy the entire instruction queue, bringing the fast thread to a halt for lack of space in the

SEC. 8.1

561

ON-CHIP PARALELLISM

instruction queue. Full resource sharing solves the problem of a resource lying idle while another thread wants it, but creates a new problem of one thread potentially hogging so many resources as to slow the other one down or stop it altogether. An intermediate scheme is threshold sharing, in which a thread can acquire resources dynamically (no fixed partitioning) but only up to some maximum. For resources that are replicated, this approach allows flexibility without the danger that one thread will starve due to its inability to acquire any of the resource. If, for example, no thread can acquire more than 3/4 of the instruction queue, no matter what the slow thread does, the fast thread will be able to run. The Pentium 4 hyperthreading uses different sharing strategies for different resources in an attempt to address the various problems alluded to above. Duplication is used for resources that each thread requires all the time, such as the program counter, register map, and interrupt controller. Duplicating these resources increases the chip area by only 5%, a modest price to pay for multithreading. Resources available in such large abundance that there is no danger of one thread capturing them all, such as cache lines, are fully shared in a dynamic way. On the other hand, resources that control the operation of the pipeline, such as the various queues within the pipeline, are partitioned, giving each thread half of the slots. The main pipeline of the Netburst microarchitecture used in the Pentium 4 is illustrated in Fig. 8-9, with the white and gray boxes indicating how the resources are allocated between the white and gray threads. PC

Trace cache

Fetch queue

Allocate/ renaming

Micro-op Scheduler Registers Execution queue

D-cache

Register write

Retirement queue

Figure 8-9. Resource sharing between threads in the Pentium 4 NetBurst microarchitecture.

In this figure we can see that all the queues are partitioned, with half the slots in each queue reserved for each thread. In this one, neither thread can choke off the other one. The register allocator and renamer is also partitioned. The scheduler is dynamically shared, but with a threshold, to prevent either thread from claiming all of the slots. The remaining pipeline stages are fully shared. All is not sweetness and light with multithreading, however. There is also a downside. While partitioning is cheap, dynamic sharing of any resource and especially with a limit on how much a thread can take, requires bookkeeping at run

562

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

time to monitor usage. In addition, situations can arise in which programs work much worse with multithreading than without it. For example, imagine two threads that each need 3/4 of the cache to function well. Run separately, each one works fine and has few (expensive) cache misses. Run together, each one has large numbers of cache misses and the net result may be far worse than without multithreading. More information about multithreading in the Pentium 4 is given in (Gerber and Binstock, 2004; Koufaty and Marr, 2003; and Tuck and Tullsen, 2003).

8.1.3 Single-Chip Multiprocessors While multithreading provides significant performance gains at modest cost, for some applications a much larger performance gain is needed than multithreading can provide. To get this gain, multiprocessor chips are being developed. Two areas where these chips, which contain two or more CPUs, are of interest are high-end servers and in consumer electronics. We will now briefly touch on each of them. Homogeneous Multiprocessors on a Chip With advances in VLSI technology, it is now possible to put two or more powerful CPUs on a single chip. Since these CPUs always share the same level 1 cache, level 2 cache, and main memory, they qualify as a multiprocessor, as discussed in Chap. 2. A typical application area is a large Web server farm consisting of many servers. By putting two CPUs in the same box, sharing not only memory, but also disks and network interfaces, the performance of the server can often be doubled without doubling the cost (because even at twice the price, the CPU chip is only a fraction of the total system cost). For small-scale single-chip multiprocessors, two designs are prevalent. In the first one, shown in Fig. 8-10(a), there is really only one chip, but it has a second pipeline, potentially doubling the instruction execution rate. In the second one, shown in Fig. 8-10(b), there are separate cores on the chip, each one containing a full CPU. A core is large circuit, such as a CPU, I/O controller, or cache, that can be placed on a chip in a modular way, usually next to other cores. The former design allows resources, such as functional units, to be shared between the processors, thus allowing one CPU to use resources the other does not need. On the other hand, this approach requires redesigning the chip and it does not scale well much above two CPUs. In contrast, just putting two or more CPU cores on the same chip is relatively easy to do. We will discuss multiprocessors later in the chapter. While that discussion is somewhat focused on multiprocessors built from single-CPU chips, much of it is also applicable to multi-CPU chips as well.

SEC. 8.1

563

ON-CHIP PARALELLISM

Pipeline

CPU

CPU

Pipeline

Cache memory

Cache memory

(a)

(b)

Figure 8-10. Single-chip multiprocessors. (a) A dual-pipeline chip. (b) A chip with two cores.

Heterogeneous Multiprocessors on a Chip A completely different application area calling for single-chip multiprocessors is embedded systems, especially in audio-visual consumer electronics, such as television sets, DVD players, camcorders, game consoles, cell phones, and so on. These systems have demanding performance requirements and tight constraints. Although these devices look different, in reality, more and more of them are simply small computers, with one or more CPUs, memories, I/O controllers, and some custom I/O devices. A cell phone, for example, is merely a PC with a CPU, memory, dwarf keyboard, microphone, loudspeaker, and a wireless network connection in a small package. Consider, as an example, a portable DVD player. The computer inside of it has to handle the following functions: 1. Control of a cheap, unreliable servomechanism for head tracking. 2. Analog to digital conversion. 3. Error correction. 4. Decryption and digital rights management. 5. MPEG-2 video decompression. 6. Audio decompression. 7. Encoding the output for NTSC, PAL, or SECAM television sets. This work must be done subject to stringent real time, quality of service, energy, heat dissipation, size, weight, and price constraints. DVD disks contain a long spiral containing the information, as illustrated in Fig. 2-24 (for a CD). The read head must accurately track the spiral as the disk rotates. The price is kept low by using a relatively simple mechanical design and

564

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

tight control over the head position in software. The signal coming off the head is an analog signal, which must be converted to digital form before being processed. After it has been digitized, heavy error correction is required because DVDs are pressed and contain many errors, which must corrected in software. The video is compressed using the MPEG-2 standard, which requires complex (Fourier transform like) computations for decompression. Audio is compressed using a psycho-acoustic model, which also requires sophisticated calculations for decompression. Finally, audio and video have to be rendered in a suitable form for output to NTSC, PAL, or SECAM television sets, depending on the country to which the DVD player is shipped. It should come as no surprise that doing all this work in real time in software on a cheap general-purpose CPU is not possible. What is needed is a heterogeneous multiprocessor containing multiple cores, each specialized for one particular task. An example DVD player is given in Fig. 8-11.

MPEG video decoder Control processor Heterogeneous multiprocessor chip with six cores

NTSC/PAL/SECAM composite video Audio encoder decoder Disk controller Cache Memory

Bus Figure 8-11. The logical structure of a simple DVD player contains a heterogeneous multiprocessor containing multiple cores for different functions.

The functions of the cores in Fig. 8-11 are all different, with each one being carefully designed to be extremely good at what it does for the lowest possible price. For example, DVD video is compressed using a scheme known as MPEG-2 (after the Motion Picture Experts Group that invented it). It consists of dividing each frame up into blocks of pixels and doing a complex transformation on each one. A frame can consist entirely of transformed blocks or it can specify that a certain block is the same as one found in the previous frame but located at an offset of (∆x, ∆y) from its current position except with a couple of pixels changed. Doing this calculation in software is extremely slow, but it is possible to build an MPEG-2 decoding engine that can do it in hardware quite rapidly. Similarly, audio decoding and reencoding the composite audio-video signal to conform to one of the world’s television standards can be done better by dedicated hardware processors. These observations quickly lead to heterogeneous multiprocessor chips containing multiple cores specifically designed for audio-

SEC. 8.1

ON-CHIP PARALELLISM

565

visual applications. However, because the control processor is a general-purpose programmable CPU, the multiprocessor chip can also be used in other, similar applications, such as a DVD recorder. Another device requiring a heterogeneous multiprocessor is the engine inside an advanced cell phone. Current ones sometimes have still cameras, video cameras, game machines, web browsers, e-mail readers, and digital satellite radio receivers, using either cell phone technology (CDMA or GSM) or wireless Internet (IEEE 802.11, also called WiFi) built in; future ones may include all of these. As devices take on more and more functionality, with watches becoming GPSbased maps and eyeglasses becoming radios, the need for heterogeneous multiprocessors will only increase. Fairly soon chips will have 500 million transistors. Such chips are far too large to design one gate and one wire at a time. The human effort required would render the chips obsolete by the time they were finished. The only feasible way is to use cores (essentially libraries) containing fairly large subassemblies and then to place and interconnect them on the chip as needed. Designers then have to determine which CPU core to use for the control processor and which specialpurpose processors to throw in to help it. Putting more of the burden on software running on the control processor makes the system slower but yields a smaller (and cheaper) chip. Having multiple special-purpose processors for audio and video processing takes up chip area, increasing the cost, but produces higher performance at a lower clock rate, which means lower power consumption and less heat dissipation. Thus chip designers are increasingly faced with these macroscopic trade-offs rather than worrying about where to place each transistor. Audio-visual applications are very data intensive. Huge amounts of data have to be processed quickly, so typically 50% to 75% of the chip area is devoted to memory in one form or another and the amount is rising. The design issues here are numerous. How many levels of cache should be used? Should the cache(s) be split or unified? How big should each cache be? How fast should each one be? Should some actual memory go on the chip too? Should it be SRAM or SDRAM? The answers to each of these questions have major implications for the performance, energy consumption, and heat dissipation of the chip. Besides design of the processors and memory system, another issue of considerable consequence is the communication system—how do all the cores communicate with each other? For small systems, a single bus will usually do the trick, but for larger ones it rapidly becomes a bottleneck. Often the problem can be solved by going to multiple buses or possibly a ring going from core to core. In the latter case, arbitration is handled by passing a small packet called a token around the ring. To transmit, a core must first capture the token. When it is done, it puts the token back on the ring so it can continue circulating. This protocol prevents collisions on the ring. As an example of an on-chip interconnect, look at the IBM CoreConnect, illustrated in Fig. 8-12. It is an architecture for connecting cores on a single-chip

566

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

heterogeneous multiprocessor, especially complete system-on-a-chip designs. In a sense, CoreConnect is to one-chip multiprocessors what the PCI bus is to the Pentium—the glue that holds all the parts together. However, unlike the PCI bus, CoreConnect was designed without any requirements to be backward compatible with legacy equipment or protocols and without the constraints of board-level buses, such as limits on the number of pins the edge connector can have.

Control CPU

Other CPU

Fast I/O Device

Device register bus

I/O device

I/O device

Arbiter

Bus bridge Peripheral bus

Processor bus Memory

Figure 8-12. An example of the IBM CoreConnect architecture.

CoreConnect consists of three buses. The processor bus is a high-speed, synchronous, pipelined bus with 32, 64, or 128 data lines clocked at 66, 133, or 183 MHz. The maximum throughput is thus 23.4 Gbps (vs. 4.2 Gbps for the PCI bus). The pipelining features allow cores to request the bus while a transfer is going on, and to allow different cores to use different lines at the same time, similar to the PCI bus. The processor bus is optimized for short block transfers. It is intended to connect fast cores, such as CPUs, MPEG-2 decoders, high-speed networks, and similar items. Stretching the processor bus over the entire chip would reduce its performance, so a second bus is present for low-speed I/O devices, such as UARTs, timers, USB controllers, serial I/O devices, and so forth. This peripheral bus has been designed with the goal of making it simple to interface 8-, 16-, and 32-bit peripherals to it using no more than a few hundred gates. It, too, is a synchronous bus, with a maximum throughput is 300 Mbps. The two buses are connected by a bridge, not unlike the bridges that were used to connect the PCI and ISA buses in PCs until the ISA bus was phased out a number of years ago. The third bus is the device register bus, a very low-speed, asynchronous, handshaking bus used to allow the processors to access the device registers of all the peripherals in order to control the corresponding devices. It is intended for infrequent transfers of only a few bytes at a time.

SEC. 8.1

ON-CHIP PARALELLISM

567

By providing a standard on-chip bus, interface, and framework, IBM hopes to create a miniature version of the PCI world, in which many manufacturers produce processors and controllers that plug together easily. One difference, however, is that in the PCI world the manufacturers produce and sell actual boards that PC vendors and end users buy. In the CoreConnect world, third parties design cores, but do not manufacture them. Instead, they license them as intellectual property to consumer electronics and other companies, which then design custom heterogeneous multiprocessor chips based on their own and licensed third-party cores. Since manufacturing such large and complex chips requires a massive investment in fabrication facilities, in most cases the consumer electronics company just does the design, subcontracting the chip manufacturing out to a semiconductor vendor. Cores for numerous CPUs (ARM, MIPS, PowerPC, etc.) exist, as well as for MPEG decoders, digital signal processors, and all the standard I/O controllers. The IBM CoreConnect is not the only popular on-chip bus on the market. The AMBA (Advanced Microcontroller Bus Architecture), is also widely used (Flynn, 1997). Other, somewhat less popular on-chip buses are the VCI (Virtual Component Interconnect) and OCP-IP (Open Core Protocol-International Partnership), which are also competing for market share (Kogel and Meyr, 2004; and Ouadjaout and Houzet, 2004). On-chip buses are only the start; people are now even thinking of complete networks on a chip (Benini and De Micheli, 2002). With chip manufacturers having increasing difficulty in raising clock frequencies due to heat dissipation problems, single-chip multiprocessors are a very hot topic. More information can be found in (Claasen, 2003; Jerraya and Wolf, 2005; Kumar et al., 2004; Lavagno, 2002; Lines, 2004; and Ravikumar, 2004).

8.2 COPROCESSORS Having examined some of the ways of achieving on-chip parallelism, let us now move up a step and look at how the computer can be speeded up by adding a second, specialized processor. These coprocessors some in many varieties, from small to large. On the IBM 360 mainframes and all of its successors, independent I/O channels exist for doing input/output. Similarly, the CDC 6600 had 10 independent processors for doing I/O. Graphics and floating-point arithmetic are other areas where coprocessors have been used. Even a DMA chip can be seen as a coprocessor. In some cases, the CPU gives the coprocessor an instruction or set of instructions and tells it to execute them; in other cases, the coprocessor is more independent and runs pretty much on its own. Physically, coprocessors can range from a separate cabinet (the 360 I/O channels) to a plug-in board (network processors) to an area on the main chip (floating-point). In all cases, what distinguishes them is the fact that some other

568

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

processor is the main processor and the coprocessors are there to help it. We will now examine thre areas where speedups are possible: network processing, multimedia, and cryptography.

8.2.1 Network Processors Most computers nowadays are connected to a network or to the Internet. As a result of technological progress in network hardware, networks are now so fast that it has become increasingly difficult to process all the incoming and outgoing data in software. As a consequence, special network processors have been developed to handle the traffic, and many high-end computers now have one of these processors. In this section we will first give a brief introduction to networking and then discuss how network processors work. Introduction to Networking Computer networks come in two general types: local-area networks, or LANs, which connect multiple computers within a building or campus, and wide-area networks or WANs, which connect computers spread over a large geographic area. The most popular LAN is called Ethernet. The original Ethernet consisted of a fat cable into which a wire coming from each computer was forcibly inserted using what was euphemistically referred to a vampire tap. Modern Ethernets have the computers attached to a central switch, as illustrated in the right-hand portion of Fig. 8-13. The original Ethernet crawled along at 3 Mbps, but the first commercial version was 10 Mbps. It was eventually replaced by fast Ethernet at 100 Mbps and then by gigabit Ethernet at 1 Gbps. A 10-gigabit Ethernet is already on the market and a 40-gigabit Ethernet is in the pipeline. Application provider premises Internet

Server

Router Firewall

Telephone line Packet ISP

Switch User’s computer

Fiber optic link

Figure 8-13. How users are connected to servers on the Internet.

SEC. 8.2

COPROCESSORS

569

WANs are organized differently. They consist of specialized computers called routers connected by wires or optical fibers, as shown in the middle of Fig. 8-13. Chunks of data called packets, typically 64 to about 1500 bytes, are moved from the source machine through one or more routers until they reach their destination. At each hop, a packet is stored in the router’s memory and then forwarded to the next router along the path as soon as the needed transmission line is available. This technique is called store-and-forward packet switching. Although many people think of the Internet as a single WAN, technically it is a collection of many WANs connected together. However, for our purposes, that distinction is not important. Figure 8-13 gives a bird’s eye view of the Internet from the perspective of a home user. The user’s computer is typically connected to a Web server via the telephone system, either a 56 kbps dial-up modem or ADSL, which was discussed in Chap. 2. (Alternatively, cable TV can be used, in which case the left-hand part of Fig. 8-13 is slightly different and the cable company is the ISP.) The user’s computer breaks the data to be sent to the server into packets and sends these packets to the user’s ISP (Internet Service Provider), a company that offers Internet access to its customers. The ISP has a high-speed (usually fiber optic) connection to one of the regional or backbone networks that comprise the Internet. The user’s packets are forwarded hop-by-hop across the Internet until they arrive at the Web server. Most companies offering Web service have a specialized computer called a firewall that filters all incoming traffic in an attempt to remove unwanted packets (e.g., from hackers trying to break in). The firewall is connected to the local LAN, typically an Ethernet switch, which routes packets to the desired server. Of course, reality is a lot more complicated than we have shown, but the basic idea of Fig. 8-13 is still valid. Network software consists of multiple protocols, each one being a set of formats, exchange sequences, and rules about what the packets mean. For example, when a user wants to fetch a Web page from a server, the user’s browser sends a packet containing a GET PAGE request using the HTTP (HyperText Transfer Protocol) to the server, which understands how to process such requests. Many protocols are in use and often combined. In most situations, protocols are structured as a series of layers, with upper layers handing packets to lower layers for processing, with the bottom layer doing the actual transmission. At the receiving side, the packets work their way up the layers in the reverse order. Since protocol processing is what network processors do for a living, it is necessary to explain a little bit about protocols before looking at the network processors themselves. Let us go back to the GET PAGE request for a moment. How is that sent to the Web server? What happens is that the browser first establishes a connection to the Web server using a protocol called TCP (Transmission Control Protocol). The software that implements this protocol checks that all packets have been correctly received and in the proper order. If a packet gets lost, the TCP software assures that it is retransmitted as often as need be until it received.

570

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

In practice, what happens is that the Web browser formats the GET PAGE request as a correct HTTP message and then hands it to the TCP software to transmit over the connection. The TCP software adds a header in front of the message containing a sequence number and other information. This header is naturally called the TCP header. When it is done, the TCP software takes the TCP header and payload (containing the GET PAGE request) and passes it to another piece of software that implements the IP protocol (Internet Protocol). This software attaches an IP header to the front containing the source address (the machine the packet is coming from), the destination address (the machine the packet is supposed to go to), how many more hops the packet may live (to prevent lost packets from living forever), a checksum (to detect transmission and memory errors), and other fields. Next the resulting packet (now consisting of the IP header, TCP header, and GET PAGE request), is passed down to the data link layer. where a data link header is attached to the front for actual transmission. The data link layer also adds a checksum to the end called a CRC (Cyclic Redundancy Code) used to detect transmission errors. It might seem that having checksums in the data link layer and the IP layer is redundant, but it improves reliability. At each hop, the CRC is checked and the header and CRC stripped and regenerated, with a format being chosen that is appropriate for the outgoing link. Figure 8-14 shows what the packet looks like when on the Ethernet; on the telephone line, it is similar except with a ‘‘telephone line header’’ instead of an Ethernet header. Header management is important and is one of the things network processors can do. Needless to say, we have only scratched the surface of the subject of computer networking. For a more comprehensive treatment, see (Tanenbaum, 2003). Ethernet header

IP header

TCP header

Payload

C R C

Figure 8-14. A packet as it appears on the Ethernet.

Introduction to Network Processors Many kinds of devices are connected to networks. End users have personal computers (desktop and notebook), of course, but increasingly also game machines, PDAs (palmtops), and cell phones. Companies have PCs and servers as end systems. However, there are also numerous devices that function as intermediate systems in networks, including routers, switches, firewalls, Web proxies, and load balancers. Interestingly enough, it is the intermediate systems that are the most demanding, since they are expected to move the largest number of packets per second. Servers are also demanding but the user machines are not.

SEC. 8.2

COPROCESSORS

571

Depending on the network and the packet itself, an incoming packet may need various kinds of processing done to it before being forwarded to either the outgoing line or the application program. This processing may include deciding where to send the packet, fragmenting it or reassembling its pieces, managing its quality of service (especially for audio and video streams), managing security (e.g., encryption or decryption), compression/decompression, and so on. With LAN speeds approaching 40 gigabits/sec and 1 KB packets, a networked computer might have to process almost 5 million packets/sec. With 64-byte packets, the number of packets that have to be processed per second rises to nearly 80 million. Performing the various functions mentioned above in 12–200 nsec (in addition to the multiple copies of the packet that are invariably needed) is just not doable in software. Hardware assistance is essential. One kind of hardware solution for fast packet processing is to use a custom ASIC (Application-Specific Integrated Circuit). Such a chip is like a hardwired program that does whatever set of processing functions it was designed for. Many current routers uses ASICs. ASICs have many problems, however. First, they take a long time to design and a long time to manufacture. They are also rigid, so if new functionality is needed, a new chip has to be designed and manufactured. Furthermore, bug management is a nightmare, since the only way to fix one is to design, manufacture, ship, and install new chips. They are also expensive unless the volume is so large as to allow amortizing the development effort over a substantial number of chips. A second solution is the FPGA (Field Programmable Gate Array) which is a collection of gates that can be organized into the desired circuit by rewiring them in the field. These chips have much a shorter time to market than ASICs and can be rewired in the field by removing them from the system and inserting them into a special reprogramming device. On the other hand, they are complex, slow, and expensive, making them unattractive except for niche applications. Finally, we come to network processors, programmable devices that can handle incoming and outgoing packets at wire speed (i.e., in real time). A common design is a plug-in board containing a network processor on a chip along with memory and support logic. One or more network lines connect to the board and are routed to the network processor. There packets are extracted, processed, and either sent out on a different network line (e.g., for a router) or are sent out onto the main system bus (e.g., the PCI bus) in the case of end-user device such as a PC. A typical network processor board and chip are illustrated in Fig. 8-15. Both SRAM and SDRAM are provided on the board and typically used in different ways. SRAM is faster, but more expensive, than SDRAM, so there is only a small amount of it. SRAM is used to hold routing tables and other key data structures, whereas SDRAM holds the actual packets being processed. By making the SRAM and SDRAM external to the network processor chip, the board designers are given the flexibility to determine how much of each to supply. In this way, low-end boards with a single network line (e.g., for a PC or server) can

572

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Network processor board SRAM

SDRAM

SRAM interface

SDRAM interface

PPE

PPE

PPE

Network processor Control CPU

PPE

Local Specialized procs. mem ...

PPE

PPE

PPE

Network interface

PPE

Buses

PCI interf. PCI connector

Figure 8-15. A typical network processor board and chip.

be equipped with a small amount of memory whereas a high-end board for a large router can be equipped with much more. Network processor chips are optimized for processing large numbers of incoming and outgoing packets quickly. Millions of packets per second per network line, that is, and a router could easily have half a dozen lines. The only way to achieve such processing rates is to build network processors that are highly parallel inside. And indeed, all network processors consist of multiple PPEs, variously called Protocol/Programmable/Packet Processing Engines. Each one is a (possibly modified) RISC core and a small amount of internal memory for holding the program and some variables. The PPEs can be organized in two different ways. The simplest organization is having all the PPEs being identical. When a packet arrives at the network processor, either an incoming packet from a network line or an outgoing packet from the bus, it is handed to an idle PPE for processing. If no PPE is idle, the packet is queued in the on-board SDRAM until a PPE frees up. When this organization is used, the horizontal connections shown between the PPEs in Fig. 8-15 do not exist because the PPEs have no need to communicate with one another. The other PPE organization is the pipeline. In this one, each PPE performs one processing step and then feeds a pointer to its output packet to the next PPE in the pipeline. In this way, the PPE pipeline acts very much like the CPU pipelines we studied in Chap. 2. In both organizations, the PPEs are completely programmable. In advanced designs, the PPEs have multithreading, meaning that they have multiple register sets and a hardware register indicating which one is currently in use. This feature is used to run multiple programs at the same time by allowing a

SEC. 8.2

COPROCESSORS

573

program (i.e., thread) switch by just changing the ‘‘current register set’’ variable. Most commonly, when a PPE stalls, for example, when it references the SDRAM (which takes multiple clock cycles), it can instantaneously switch to a runnable thread. In this manner, a PPE can achieve a high utilization even when frequently blocking to access the SDRAM or perform some other slow external operation. In addition to the PPEs, all network processors contain a control processor, usually just a standard general-purpose RISC CPU, for performing all work not related to packet processing, such as updating the routing tables. Its program and data are in the local on-chip memory. Furthermore, many network processor chips also contain one or more specialized processors for doing pattern matching or other critical operations. These processors are really small ASICs that are good at doing one simple operation, such as looking up a destination address in the routing table. All the components of the network processor communicate over one or more on-chip, parallel buses that run at multigigabit/sec speeds. Packet Processing When a packet arrives, it goes through a number of processing stages, independent of whether the network processor has a parallel or pipeline organization. Some network processors divide these steps into operations performed on incoming packets (either from a network line or from the system bus), called ingress processing and operations performed on outgoing packets, called egress processing. When this distinction is made, every packet first goes through ingress processing, then through egress processing. The boundary between ingress and egress processing is flexible because some steps can be done in either part (e.g., collecting traffic statistics). Below we will discuss a potential ordering of the various steps, but note that not all packets need all steps and that many other orderings are equally valid. 1. Checksum verification. If the incoming packet is arriving from the Ethernet, the CRC is recomputed so it can be compared with the one in the packet to make sure there was no transmission error. If the Ethernet CRC is correct or not present, the IP checksum is recomputed and compared to the one in the packet to make sure the IP packet was not damaged by a faulty bit in the sender’s memory after the IP checksum was computed there. If all checksums are correct, the packet is accepted for further processing; otherwise, it is simply discarded. 2. Field extraction. The relevant header is parsed and key fields are extracted. In an Ethernet switch, only the Ethernet header is examined, whereas in an IP router, it is the IP header that is inspected. The key fields are stored in registers (parallel PPE organization) or SRAM (pipeline organization).

574

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

3. Packet classification. The packet is classified according to a series of programmable rules. The simplest classification is to distinguish data packets from control packets, but usually much finer distinctions are made. 4. Path selection. Most network processors have a special fast path optimized for handling plain old garden-variety data packets, with all other packets being treated differently, often by the control processor. Consequently, either the fast path or the slow path has to be selected. 5. Destination network determination. IP packets contain a 32-bit destination address. It is not possible (or desirable) to have a 232 entry table to lookup the destination of each IP packet, so the leftmost part of each IP address is the network number and the rest specifies a machine on that network. Network numbers can be of any length, so determining the destination network number is nontrivial and made worse by the fact that multiple matches are possible and the longest one counts. Often a custom ASIC is used in this step. 6. Route lookup. Once the number of the destination network is known, the outgoing line to use can be looked up in a table in the SRAM. Again, a custom ASIC may be used in this step. 7. Fragmentation and reassembly. Programs like to present large payloads to the TCP layer to reduce the number of system calls needed, but TCP, IP, and Ethernet all have maximum sizes for the packets they can handle. As a consequence of these limits, payloads and packets may have to be fragmented at the sending side and the pieces reassembled at the receiving side. These are tasks the network processor can perform. 8. Computation. Heavy-duty computation on the payload is sometimes required, for example, data compression/decompression and encryption/decryption. These are tasks a network processor can perform. 9. Header management. Sometimes headers have to be added, removed, or have some of their fields modified. For example, the IP header has a field that counts the number of hops the packet may yet make before being discarded. Every time it is retransmitted, this field must be decremented, something the network processor can do. 10. Queue management. Incoming and outgoing packets often have to be queued while waiting their turn at being processed. Multimedia applications may need a certain interpacket spacing in time to avoid

SEC. 8.2

COPROCESSORS

575

jitter. A firewall or router may need to distribute the incoming load among multiple outgoing lines according to certain rules. All of these tasks can be done by the network processor. 11. Checksum generation. Outgoing packets need to be checksummed. The IP checksum can be generated by the network processor, but the Ethernet CRC is generally computed by hardware. 12. Accounting. In some cases, accounting for packet traffic is needed, especially when one network is forwarding traffic for other networks as a commercial service. The network processor can do the accounting. 13. Statistics gathering. Finally, many organizations like to collect statistics about their traffic, and the network processor is a good place to collect them. Improving Performance Performance is the name of the game for network processors. What can be done to improve it? But before improving it, we have to define what it means. One metric is the number of packets that are forwarded per second. A second one is the number of bytes forwarded per second. These are different measures and a scheme that works well with small packets may not work as well with large ones. In particular, with small packets, improving the number of destination lookups per second may help a lot, but with large packets it may not. The most straightforward way to improve performance is to increase the speed of the network processor clock. Of course, performance is not linear with clock speed, since memory cycle time and other factors also influence it. Also, a faster clock means more heat must be dissipated. Introducing more PPEs and parallelism is often a winner, especially with an organization consisting of parallel PPEs. A deeper pipeline can also help, but only if the job of processing a packet can be split up into smaller pieces. Another technique is to add specialized processors or ASICs to handle specific, time consuming operations that are performed repeatedly and that can be done faster in hardware than in software. Lookups, checksum computations, and cryptography are among the many candidates. Adding more internal buses and widening existing buses may help gain speed by moving packets through the system faster. Finally, replacing SDRAM by SRAM can usually be counted to improve performance, but at a price, of course. There is much more that can be said about network processors, of course. Some references are (Comer, 2005; Crowley et al., 2002; Lekkas, 2003; and Papaefstathiou et al., 2004).

576

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

8.2.2 Media Processors A second area in which coprocessors are used is for handling high-resolution photographic images, audio, and video streams. Ordinary CPUs are not especially good at the massive computations needed to process the large amounts of data required in these applications. For this reason, some current PCs and most future PCs will be equipped with media coprocessors to which they can offload large portions of the work. The Nexperia Media Processor We will study this increasingly important area by means of an example: the Philips Nexperia, a family of chips that is available at several clock frequencies. The Nexperia is a self-contained single-chip heterogeneous multiprocessor in the sense of Fig. 8-11. It contains multiple cores, including a TriMedia VLIW CPU for control, but also numerous cores for image, audio, video, and networking processing. It can be used either as a stand-alone main processor in a CD, DVD, or MP3 player or recorder, TV set or set-top box, still or video camera, etc., or as a coprocessor in a PC for processing images and media streams. In both configurations, it runs its own small real-time operating system. The Nexperia has three functions: capturing input streams and converting them to data structures in memory, processing these data structures, and finally, outputting them in forms suitable for the various output devices attached. For example, when a PC is used as a DVD player, the Nexperia can be programmed to read the encrypted, compressed video stream from the DVD disc, decrypt and decompress it, and then output it at a size appropriate to the window it is being displayed in. And all of this can be done in the background, without involving the computer’s main CPU at all once the DVD player program has been loaded into the Nexperia. All incoming data are first stored in memory for processing; there is no direct connection between input devices and output devices. Capturing input includes decoding from a wide variety of video sizes and formats (including MPEG-1, MPEG-2, and MPEG-4), audio formats (including AAC, Dolby, and MP3) and converting to appropriate data structures for storage and processing. Input can come from the PCI bus, Ethernet, or dedicated input lines (e.g., a microphone or stereo system plugged directly into the chip). The Nexperia chip has 456 pins, some of which are available for direct input and output of media (and other) streams. Data processing is controlled by the software in the TriMedia CPU, which can be programmed for whatever is needed. Typical tasks include deinterlacing video to improve its sharpness, correcting the brightness, contrast, and color of images, scaling image size, converting between different video formats, and reducing

SEC. 8.2

577

COPROCESSORS

noise. Usually, the CPU acts as the general contractor, subcontracting out much of the work to the specialized cores on the chip. Output functionality includes coding the data structures in a form suitable for the output device, merging multiple (video, audio, image, 2D-graphics) data sources, and controlling the output devices. As with input, output can go to the PCI bus, Ethernet, or dedicated output lines (e.g., a loudspeaker or amplifier). A block diagram of the Nexperia PNX 1500 chip is given in Fig. 8-16. Other versions are slightly different, so to be consistent, throughout this section when we say ‘‘Nexperia’’ we mean the PNX 1500 implementation. It has four major sections: control, input, processing, and output. The CPU is the 32-bit TriMedia VLIW processor discussed in Sec. 8.1.1 running at 300 MHz. Its program, usually written in C or C++, determines the Nexperia’s functionality. Control

Input TriMedia Memory interface

To PCI bus

PCI bus interface

To Ethernet

Ethernet

Processing

Boot, reset, clocks

DVD descrambler Length decoder

Graphics engine JTAG

Debugging

32

8 Audio channels

Audio input SP digital Input 16

Misc., I/O, timers

Output Composite video Generic output

Router

Video scaler

Generic input

On-chip system bus

27 MHz crystal

Video input

Router

To SDRAM

Audio output

32

8 Audio channels

SP digital output

Figure 8-16. The Nexperia heterogeneous multiprocessor on a chip.

The Nexperia does not contain any main memory on chip, except for the two caches inside the TriMedia CPU. Instead it has an interface to external memory, allowing for 8–256 MB of DDR SDRAM, which is plenty for typical multimedia applications. With a clock of 200 MHz, the memory bandwidth is 1.6 GB/sec.

578

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

A full PCI interface is also included on chip, with 8-, 16-, and 32-bit transfers at 33 MHz. When used as the main CPU inside a consumer electronics device (e.g., a DVD player), the PCI interface can also act as bus arbiter. This interface can be used, for example, to talk to a DVD drive. Direct ethernet connectivity is provided by a dedicated core that can handle 10- and 100-Mbps Ethernet connections. Consequently, a Nexperia-based camcorder can output a digital video stream over an Ethernet to a remote capture or display device. The next core handles booting, resetting, clocks, and some other minor features. If a certain pin on the Nexperia is asserted, a reset is initiated. The core can also be programmed as a dead man’s switch. If the CPU fails to ping it for a certain period of time, it assumes the system has hung and initiates a reboot on its own. In stand-alone devices, rebooting can be done from a flash memory. The core is driven by an external 27 MHz crystal oscillator, which is multiplied internally by 64 to give a 1.728 GHz signal used throughout the chip. Power management is also handled here. Normally, the CPU runs at full speed and the other components run at whatever speed they have to to get their work done. However, it is possible for the CPU to slow down the clock to save energy. A sleep mode is also provided to shut down most functions when there is no work to do, thus conserving battery charge on mobile devices. This unit also contains 16 ‘‘semaphores’’ that can be used for interdevice synchronization. When a core writes a nonzero value to a semaphore that is currently zero, the write succeeds; otherwise, it fails and the old value is left intact. Writing a zero always succeeds. Since only one core can acquire the system bus at a time, these operations are atomic and provide the necessary tools for achieving mutual exclusion. To acquire a resource, a core attempts to write a nonzero value to a specific semaphore. If the write succeeds, it has exclusive access to the resource until it releases the resource by writing a zero to it. When a write fails, the core that failed has to keep trying periodically until it succeeds, so these are not really classical semaphores in the sense of Chap. 6. Now let us look at the input section. The video input core takes an incoming 10-bit wide digital video stream, converts it to 8 bits using a smoothing algorithm, and stores it in the external SDRAM. In most situations, the digital input is the output of an external analog-to-digital converter that is being fed an analog television signal, either over the air or from a cable TV input. The generic input core can capture unstructured 32-bit wide input from a arbitrary data source at 100 MHz and store it in the SDRAM. It can also capture structured data with markers delineating the record boundaries. The router in front of the two digital video inputs demultiplexes them and can also do some video transformations on the fly. Demultiplexing is needed because the same pins are used for both video input and generic input. The audio input core can capture up to eight channels of stereo music or voice with 8-, 16-, or 32-bit precision at up to 96 kHz and store it in the SDRAM. It can

SEC. 8.2

COPROCESSORS

579

also decode compressed formats, mix channels, change the sampling rate, and apply filters, all on the fly before storing the audio data. The SP digital input core allows inputing digital audio signals that conform to the Sony-Philips digital audio standard (IEC 1937). In this way, digital audio can be moved from one device to another with no loss in quality. Once the audio, video, or other data has been input, it normally needs to be processed, which is what the next section handles. Movies sold or rented on DVDs are encrypted (scrambled) to prevent copying. The DVD descrambler removes the encryption to provide the original movie, compressed with MPEG-2. This decryption is a memory-to-memory operation, with the input coming from one buffer and the output going to a different buffer. The length decoder goes further and removes the variable length coding features of MPEG-2 (or MPEG-1) compression producing semi-uncompressed data for the core of the MPEG processing, which is done by the TriMedia in software. The reason for this split is that the removal of the variable-length decoding (Huffman and run-length decoding) does not uses the TriMedia’s capabilities very efficiently, so it was deemed better to use a couple of square millimeters of silicon to do it in hardware. These operations, all memory to memory, produce a straightforward pixel map. The pixel map can be in any one of three general formats, each with three or four variants for different sizes and parameters. The first format is indexed color, in which each value is an index into a CLUT (Color Look Up Table) containing a 24-bit color value and an 8-bit alpha channel mask, used for determining transparency when multiple layers are overlaid. The second format is RGB, the way computer monitors work, with separate values for the red, green, and blue intensity of each pixel. The third format is YUV, designed for how television signals are encoded. Rather than encoding the red, green, and blue components separately, a transformation is done in the camera to produce a luminance (brightness) channel and two chrominance (color) channels. This system allows more bandwidth to be allocated to luminance than chrominance, which, in turn, provides better noise immunity during transmission. For an application involving television in and television out, the YUV format makes sense. By restricting the stored format to a limited number of options, each core is capable of reading the output produced by all the other cores. The video scaler accepts a list of scaling tasks and then performs them, at a maximum rate of 120 million pixels/sec. These tasks include: 1. 2. 3. 4. 5. 6.

Deinterlacing. Horizontal and vertical scaling. Linear and nonlinear aspect ratio conversion. Conversion between different pixel formats. Luminance histogram collection. Flicker reduction.

580

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Broadcast television signals are interlaced, meaning that for each frame, which consists of 525 scan lines (625 for PAL and SECAM), first all the even lines are transmitted, then the odd ones. Deinterlacing produces a better-quality progressive scan, in which all the scan lines are processed or transmitted in their true order and refreshed at twice the interlaced frame rate (29.97 fps for NTSC and 25 fps for PAL and SECAM). Horizontal and vertical scaling allow images to be increased or reduced in size, possibly after cropping. Standard television has a width:height aspect ratio of 4:3, but wide-screen television is 16:9, more suited to the 3:2 ratio of 35 mm movies. The scaler can convert between these aspect ratios, either linearly or using a nonlinear algorithm. It can also convert among the index, RGB, and YUV formats and build a histogram of the luminance value, which is useful for improving image quality on output. Finally, certain transformations can be done to reduce image flicker. The graphics engine does two-dimensional rendering from object descriptions. It can also fill enclosed areas and perform bitblt graphic operations, which amount to taking two rectangular pixel maps and combining them using AND, OR, XOR, or some other Boolean function. There are no cores for doing audio processing. All audio processing not done on input is done by the TriMedia CPU in software. Audio requires so little data that processing it in software is not a problem at all. Also, many applications do not require any audio processing at all, except possibly, format conversions. The debugging core helps the designers and programmers debug the hardware and software. It provides an interface to JTAG (Joint Test Action Group) aware instrumentation and tools, as defined in IEEE standard 1149.1. The output section takes the processed data from memory and outputs it. The composite video core takes one or more data structures representing pixels, normalizes them in certain ways and then blends them in specified ways for output. Data in indexed format is de-indexed on the fly to get the actual pixels, and incompatible formats are converted as need be. This core can also do contrast, brightness, and color correction, if required. It can also perform chroma keying, in which an actor standing in front of a pure blue screen is isolated from the screen and visually placed in front of a background taken from a different source. Similarly, animated cartoons in which foreground characters move in front of a still or scrolling background can be produced here. And, of course, the final result is converted to the video or television system format required (e.g., NTSC, PAL, or SECAM), including the generation of horizontal and vertical sync pulses. Since it costs nothing extra, it is expected that most Nexperia-based systems will be able to handle all three television formats automatically so that they can be sold unmodified anywhere in the world. Similarly, adding HDTV (High Definition TeleVision) to the mix in any of the formats means just adding a bit more software to handle conversions to and from the memory data structures. The generic output core just moves bits, 8, 16, or 32 per cycle at 100 MHz, giving a maximum bandwidth of 3.2 Gbps. By interfacing the generic output of

SEC. 8.2

COPROCESSORS

581

one Nexperia chip to the generic input of another Nexperia chip, file transfers can be performed at speeds exceeding gigabit Ethernet (1 Gbps). This interface also allows the CPU to produce custom output of whatever type it needs, in software. The output router multiplexes the two output sources and also adds some functionality, including refreshing flat panel (TFT) displays of sizes up to 1280 × 768 pixels at 60 Hz or refreshing interlaced or progressive television sets. Multiplexing is needed because the composite video and generic outputs share the same pins. The audio output core can produce up to 8 stereo channels with 32-bit precision at sampling rates up to 96 kHz. Commonly, this output drives an external digital-to-analog converter. The SP digital output can be connected to the SP digital input of devices that use the Sony-Philips digital audio standard. The final core handles general-purpose I/O. Sixteen pins are available for any use required. They can be connected to buttons, switches, or LEDs to allow them to be sensed or activated in software. They can even be used for medium speed (20 Mbps) software-controlled network protocols. Various timers, counters, and event handlers are also present here. All in all, the Nexperia has an enormous amount of computing power for audio-visual applications, and like network processors, allows the CPU to offload a massive amount of work. The amount of computing power is even larger than it may first appear, since all the cores can run in parallel with each other and with the CPU. And perhaps surprisingly, it costs less than $20 when purchased in large volume. By now the power of coprocessors, especially those based on heterogeneous multiprocessor chips, should be coming clearer. For an analogous chip aimed at telephony rather than multimedia, see (Nickolls et al., 2003).

8.2.3 Cryptoprocessors A third area in which coprocessors are popular is security, especially network security. When a connection is established between a client and a server, in many cases they must first authenticate each other. Then a secure, encrypted connection has to be established between them so data can be transferred in a secure way to foil any snoopers who may tap the line. The problem with security is that to achieve it, cryptography has to be used, and cryptography is very compute intensive. Cryptography comes in two general flavors, called symmetric key cryptography and public-key cryptography. The former is based on mixing up the bits very thoroughly, sort of the electronic equivalent of throwing a message into an electric blender. The latter is based on multiplication and exponeniation of large (e.g., 1024-bit) numbers and is extremely time consuming. To handle the computation needed to encrypt data securely for transmission or storage and then decrypt it later, various companies have produced crypto coprocessors, sometimes as PCI bus plug-in cards. These coprocessors have special

582

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

hardware that enables them to do the necessary cryptography much faster than an ordinary CPU can do it. Unfortunately, a detailed discussion of how cryptoprocessors work would first require explaining quite a bit about cryptography itself, which is beyond the scope of this book. For more information about crypto coprocessors, see (Daneshbeh and Hasan, 2004; and Lutz and Hasan, 2004).

8.3 SHARED-MEMORY MULTIPROCESSORS We have now seen how parallelism can be added to single chips and to individual systems by adding a coprocessor. The next step is to see how multiple fullblown CPUs can be combined into larger systems. Systems with multiple CPUs can be divided into multiprocessors and multicomputers. After taking a close look at what these terms actually mean, we will first study multiprocessors and then multicomputers.

8.3.1 Multiprocessors vs. Multicomputers In any parallel computer system, CPUs working on different parts of the same job must communicate with one another to exchange information. Precisely how they should do this is the subject of much debate in the architectural community. Two distinct designs, multiprocessors and multicomputers, have been proposed and implemented. The key difference between the two is the presence or absence of shared memory. This difference permeates how they are designed, built, and programmed, as well as their scale and price. Multiprocessors A parallel computer in which all the CPUs share a common memory is called a multiprocessor, as indicated symbolically in Fig. 8-17. All processes working together on a multiprocessor can share a single virtual address space mapped onto the common memory. Any process can read or write a word of memory by just executing a LOAD or STORE instruction. Nothing else is needed. The hardware does the rest. Two processes can communicate by simply having one of them write data to memory and having the other one read them back. The ability for two (or more) processes to communicate by just reading and writing memory is the reason multiprocessors are popular. It is an easy model for programmers to understand and is applicable to a wide range of problems. Consider, for example, a program that inspects a bit-map image and lists all the objects in it. One copy of the image is kept in memory, as shown in Fig. 8-17(b). Each of the 16 CPUs runs a single process, which has been assigned one of the 16 sections to analyze. Nevertheless, each process has access to the entire image, which is essential, since some objects may occupy multiple sections. If a process

SEC. 8.3

583

SHARED-MEMORY MULTIPROCESSORS P

P

P

P

P P

Shared memory

P P P

P

P (a)

P

CPU

P

P

P

P

P

P

P

P

P

P

P

P

P

P

P

P P

P

P

P

(b)

Figure 8-17. (a) A multiprocessor with 16 CPUs sharing a common memory. (b) An image partitioned into 16 sections, each being analyzed by a different CPU.

discovers that one of its objects extends over a section boundary, it just follows the object into the next section by reading the words of that section. In this example, some objects will be discovered by multiple processes, so some coordination is needed at the end to determine how many houses, trees, and airplanes there are. Because all CPUs in a multiprocessor see the same memory image, there is only one copy of the operating system. Consequently, there is only one page map and one process table. When a process blocks, its CPU saves its state in the operating system tables, and then looks in those tables to find another process to run. It is this single-system image that distinguishes a multiprocessor from a multicomputer, in which each computer has its own copy of the operating system. A multiprocessor, like all computers, must have I/O devices, such as disks, network adaptors, and other equipment. In some multiprocessor systems, only certain CPUs have access to the I/O devices, and thus have a special I/O function. In other ones, every CPU has equal access to every I/O device. When every CPU has equal access to all the memory modules and all the I/O devices, and is treated as interchangeable with the others by the operating system, the system is called an SMP (Symmetric MultiProcessor). Multicomputers The second possible design for a parallel architecture is one in which each CPU has its own private memory, accessible only to itself and not to any other CPU. Such a design is called a multicomputer, or sometimes a distributed memory system, and is illustrated in Fig. 8-18(a). The key aspect of a multicomputer that distinguishes it from a multiprocessor is that each CPU in a multicomputer has its own private, local memory that it can access by just executing LOAD and STORE instructions, but which no other CPU can access using LOAD and

584

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

STORE instructions. Thus multiprocessors have a single physical address space shared by all the CPUs whereas multicomputers have one physical address space per CPU. Since the CPUs on a multicomputer cannot communicate by just reading and writing the common memory, they need a different communication mechanism. What they do is pass messages back and forth using the interconnection network. Examples of multicomputers include the IBM BlueGene/L, Red Storm, and the Google cluster.

M

P

M

P

M

P

M

P

M

M

M

M

Private memory

P

P

P

P

CPU

Messagepassing interconnection network

P

P

P

P

M

M

M

M

(a)

P

P

M

P

P

M

P

P

M

P

P

M

P

P

P

P

CPU P

Messagepassing interconnection network

P P P

P

P

P

P

(b)

Figure 8-18. (a) A multicomputer with 16 CPUs, each with its own private memory. (b) The bit-map image of Fig. 8-17 split up among the 16 memories.

The absence of hardware shared memory on a multicomputer has important implications for the software structure. Having a single virtual address space with all processes being able to read and write all of memory by just executing LOAD and STORE instructions is impossible on a multicomputer. For example, if CPU 0 (the one in the upper left-hand corner) of Fig. 8-17(b) discovers that part of its object extends into the section assigned to CPU 1, it can nevertheless just continue reading memory to access the tail of the airplane. On the other hand, if CPU 0 in Fig. 8-18(b) makes the same discovery, it cannot just read CPU 1’s memory. It has to do something quite different to get the data it needs. In particular, it has to discover (somehow) which CPU has the data it needs and send that CPU a message requesting a copy of the data. Typically it will then block until the request is answered. When the message arrives at CPU 1, software there has to analyze it and send back the needed data. When the reply message gets back to CPU 0, the software is unblocked and can continue executing.

SEC. 8.3

SHARED-MEMORY MULTIPROCESSORS

585

On a multicomputer, communication between processes often uses software primitives such as send and receive. This gives the software a different, and far more complicated structure, than on a multiprocessor. It also means that correctly dividing up the data and placing them in the optimal locations is a major issue on a multicomputer. It is less of an issue on a multiprocessor since placement does not affect correctness or programmability although it may affect performance. In short, programming a multicomputer is much more difficult than programming a multiprocessor. Under these conditions, why would anyone build multicomputers, when multiprocessors are easier to program? The answer is simple: large multicomputers are much simpler and cheaper to build than multiprocessors with the same number of CPUs. Implementing a memory shared by even a few hundred CPUs is a substantial undertaking, whereas building a multicomputer with 10,000 CPUs or more is straightforward. Later in this chapter we will study a multicomputer with over 50,000 CPUs. Thus we have a dilemma: multiprocessors are hard to build but easy to program whereas multicomputers are easy to build but hard to program. This observation has led to a great deal of effort to construct hybrid systems that are relatively easy to build and relatively easy to program. This work has led to the realization that shared memory can be implemented in various ways, each with its own set of advantages and disadvantages. In fact, much of the research in parallel architectures these days relates to the convergence of multiprocessor and multicomputer architectures into hybrid forms that combine the strengths of each. The holy grail here is to find designs that are scalable, that is, continue to perform well as more and more CPUs are added. One approach to building hybrid systems is based on the fact that modern computer systems are not monolithic but are constructed as a series of layers—the theme of this book. This insight opens the possibility of implementing the shared memory at any one of several layers, as shown in Fig. 8-19. In Fig. 8-19(a) we see the shared memory being implemented by the hardware as a true multiprocessor. In this design, there is a single copy of the operating system with a single set of tables, in particular, the memory allocation table. When a process needs more memory, it traps to the operating system, which then looks in its table for a free page and maps the page into the caller’s address space. As far as the operating system is concerned, there is a single memory and it keeps track of which process owns which page in software. There are many ways to implement hardware shared memory, as we will see later. A second possibility is to use multicomputer hardware and have the operating system simulate shared memory by providing a single system-wide paged shared virtual address space. In this approach, called DSM (Distributed Shared Memory) (Li and Hudak, 1989), each page is located in one of the memories of Fig. 8-18(a). Each machine has its own virtual memory and its own page tables. When a CPU does a LOAD or STORE on a page it does not have, a trap to the

586

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Machine 1

Machine 2

Machine 1

Machine 2

Machine 1

Machine 2

Application

Application

Application

Application

Application

Application

Language run-time system

Language run-time system

Language run-time system

Language run-time system

Language run-time system

Language run-time system

Operating system

Operating system

Operating system

Operating system

Operating system

Operating system

Hardware

Hardware

Hardware

Hardware

Hardware

Hardware

Shared memory

Shared memory

Shared memory

(a)

(b)

(c)

Figure 8-19. Various layers where shared memory can be implemented. (a) The hardware. (b) The operating system. (c) The language runtime system.

operating system occurs. The operating system then locates the page and asks the CPU currently holding it to unmap the page and send it over the interconnection network. When it arrives, the page is mapped in and the faulting instruction is restarted. In effect, the operating system is just satisfying page faults from remote memory instead of from disk. To the user, the machine looks as if it has shared memory. We will examine DSM later in this chapter. A third possibility is to have a user-level runtime system implement a (possibly language-specific) form of shared memory. In this approach, the programming language provides some kind of shared memory abstraction, which is then implemented by the compiler and runtime system. For example, the Linda model is based on the abstraction of a shared space of tuples (data records containing a collection of fields). Processes on any machine can input a tuple from the shared tuple space or output a tuple to the shared tuple space. Because access to the tuple space is controlled entirely in software (by the Linda runtime system), no special hardware or operating system support is needed.

SEC. 8.3

SHARED-MEMORY MULTIPROCESSORS

587

Another example of a language-specific shared memory implemented by the runtime system is the Orca model of shared data objects. In Orca, processes share generic objects rather than just tuples and can execute object-specific methods on them. When a method changes the internal state of an object, it is up to the runtime system to make sure all copies of the object on all machines are simultaneously updated. Again, because objects are a strictly software concept, the implementation can be done by the runtime system without help from the operating system or hardware. We will look at both Linda and Orca later in this chapter. Taxonomy of Parallel Computers Now let us get back to our main topic, the architecture of parallel computers. Many kinds of parallel computers have been proposed and built over the years, so it is natural to ask if there is some way of categorizing them into a taxonomy. Many researchers have tried, with mixed results (Flynn, 1972; and Treleaven, 1985). Unfortunately, the Carolus Linnaeus† of parallel computing is yet to emerge. The only scheme that is used much is Flynn’s, and even his is, at best, a very crude approximation. It is given in Fig. 8-20. ######################################################################### ! Instruction ! Data ! ! ! ! ! ! ! ! streams Examples !######################################################################### ! streams ! Name ! ! !######################################################################### ! 1 ! SISD ! Classical Von Neumann machine ! 1 ! ! ! ! ! 1 !######################################################################### ! Multiple ! SIMD ! Vector supercomputer, array processor ! ! Multiple ! 1 ! MISD ! Arguably none ! !######################################################################### ! ! ! ! Multiple !######################################################################### ! Multiple ! MIMD ! Multiprocessor, multicomputer ! Figure 8-20. Flynn’s taxonomy of parallel computers.

Flynn’s classification is based on two concepts—instruction streams and data streams. An instruction stream corresponds to a program counter. A system with n CPUs has n program counters, hence n instruction streams. A data stream consists of a set of operands. The temperature computation example given earlier has multiple data streams, one for each sensor. The instruction and data streams are, to some extent, independent, so four combinations exist, as listed in Fig. 8-20. SISD is just the classical, sequential von Neumann computer. It has one instruction stream, one data stream, and does one thing at a time. SIMD machines have a single control unit that issues one instruction at a time, but they have multiple ALUs to carry it out on multiple data sets simultaneously. The ILLIAC IV (Fig. 2-7) is the prototype of SIMD """"""""""""""""

† Carolus Linnaeus (1707-1778) was the Swedish biologist who devised the system now used for classifying all plants and animals into kingdom, phylum, class, order, family, genus, and species.

588

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

machines. Mainstream SIMD machines are increasingly rare, but conventional computers sometimes have some SIMD instructions for processing audio-visual material. The Pentium SSE instructions are SIMD. Nevertheless, there is one new area in which some of the ideas from the SIMD world are playing a role: stream processors. These machines are specifically designed to handle the demands of multimedia rendering and may become important in the future (Kapasi et al., 2003). MISD machines are a somewhat strange category, with multiple instructions operating on the same piece of data. It is not clear if any such machines exist, although some people regard pipelined machines as MISD. Finally, we have MIMD, which are just multiple independent CPUs operating as part of a larger system. Most parallel processors fall into this category. Both multiprocessors and multicomputers are MIMD machines. Flynn’s taxonomy stops here, but we have extended it in Fig. 8-21. SIMD has been split into two subgroups. The first one is for numeric supercomputers and other machines that operate on vectors, performing the same operation on each vector element. The second one is for parallel-type machines, such as the ILLIAC IV, in which a master control unit broadcasts instructions to many independent ALUs.

Parallel computer architectures

SISD

SIMD

MISD

(Von Neumann) Vector processor

MIMD

? Array processor

UMA

Bus

Multiprocessors

COMA

Switched

Multicomputers

NUMA

CC-NUMA

NC-NUMA

Shared memory

Figure 8-21. A taxonomy of parallel computers.

MPP

Grid

COW

Hypercube

Message passing

SEC. 8.3

SHARED-MEMORY MULTIPROCESSORS

589

In our taxonomy, the MIMD category has been split into multiprocessors (shared-memory machines), and multicomputers (message-passing machines). Three kinds of multiprocessors exist, distinguished by the way the shared memory is implemented on them. They are called UMA (Uniform Memory Access), NUMA (NonUniform Memory Access), and COMA (Cache Only Memory Access). These categories exist because in large multiprocessors, the memory is usually split up into multiple modules. UMA machines have the property that each CPU has the same access time to every memory module. In other words, every memory word can be read as fast as every other memory word. If this is technically impossible, the fastest references are slowed down to match the slowest ones, so programmers do not see the difference. This is what ‘‘uniform’’ means here. This uniformity makes the performance predictable, an important factor for writing efficient code. In contrast, in a NUMA multiprocessor, this property does not hold. Often there is a memory module close to each CPU and accessing that memory module is much faster than accessing distant ones. The result is that for performance reasons, it matters where code and data are placed. COMA machines are also nonuniform, but in a different way. We will study each of these types and their subcategories in detail later. The other main category of MIMD machines consists of the multicomputers, which, unlike the multiprocessors, do not have shared primary memory at the architectural level. In other words, the operating system on a multicomputer CPU cannot access memory attached to a different CPU by just executing a LOAD instruction. It has to send an explicit message and wait for an answer. The ability of the operating system to read a distant word by just doing a LOAD is what distinguishes multiprocessors from multicomputers. As we mentioned before, even on a multicomputer, user programs may have the ability to access remote memory by using LOAD and STORE instructions, but this illusion is supported by the operating system, not the hardware. This difference is subtle, but very important. Because multicomputers do not have direct access to remote memory, they are sometimes called NORMA (NO Remote Memory Access) machines. Multicomputers can be roughly divided into two categories. The first category contains the MPPs (Massively Parallel Processors), which are expensive supercomputers consisting of many CPUs tightly coupled by a high-speed proprietary interconnection network. The IBM SP/3 is a well-known commercial example. The other category consists of regular PCs or workstations, possibly rackmounted, and connected by commercial off-the-shelf interconnection technology. Logically, there is not much difference, but huge supercomputers costing many millions of dollars are used differently than networks of PCs assembled by the users for a fraction of the price of an MPP. These home-brew machines go by various names, including NOW (Network of Workstations), COW (Cluster of Workstations), or sometimes just cluster.

590

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

8.3.2 Memory Semantics Even though all multiprocessors present the CPUs with the image of a single shared address space, often there are many memory modules present, each holding some portion of the physical memory. The CPUs and memories are often connected by a complex interconnection network, as discussed in Sec. 8.1.2. Several CPUs may be attempting to read a memory word at the same time several other CPUs are attempting to write the same word, and some of the request messages may pass each other in transit and be delivered in a different order than they were issued. Add to this problem the existence of multiple copies of some blocks of memory (e.g., in caches), and the result can easily be chaos unless strict measures are taken to prevent it. In this section we will see what shared memory really means and look at how memories can reasonably respond under these circumstances. One view of memory semantics is to see it as a contract between the software and the memory hardware (Adve and Hill, 1990). If the software agrees to abide by certain rules, the memory agrees to deliver certain results. The discussion then centers around what the rules are. These rules are called consistency models, and many different ones have been proposed and implemented. To give an idea of what the problem is, suppose that CPU 0 writes the value 1 to some memory word and a little later CPU 1 writes the value 2 to the same word. Now CPU 2 reads the word and gets the value 1. Should the computer owner bring the computer to the repair shop to get it fixed? That depends on what the memory promised (its contract). Strict Consistency The simplest model is strict consistency. With this model, any read to a location x always returns the value of the most recent write to x. Programmers love this model, but it is effectively impossible to implement in any way other than having a single memory module that simply services all requests first-come, first-served, with no caching and no data replication. Such an implementation would make memory an enormous bottleneck and is thus not a serious candidate, unfortunately. Sequential Consistency Next best is a model called sequential consistency (Lamport, 1979). The idea here is that in the presence of multiple read and write requests, some interleaving of all the requests is chosen by the hardware (nondeterministically), but all CPUs see the same order.

SEC. 8.3

591

SHARED-MEMORY MULTIPROCESSORS

To see what this means, consider an example. Suppose that CPU 1 writes the value 100 to word x, and 1 nsec later CPU 2 writes the value 200 to word x. Now suppose that 1 nsec after the second write was issued (but not necessarily completed yet) two other CPUs, 3 and 4, read word x twice in rapid succession, as shown in Fig. 8-22(a). Three possible orderings of the six events (two writes and four reads) are shown in Fig. 8-22(b)–(d), respectively. In Fig. 8-22(b), CPU 3 gets (200, 200) and CPU 4 gets (200, 200). In Fig. 8-22(c), they get (100, 200) and (200, 200), respectively. In Fig. 8-22(d), they get (100, 100) and (200, 100), respectively. All of these are legal, as well as some other possibilities that are not shown. CPU 2 Write 200 1

Write 100

x

Read 2x

Read 2x

3

W100

W100

W200

W200

R3 = 100

R4 = 200

R3 = 200

W200

W100

R3 = 200

R4 = 200

R3 = 100

R4 = 200

R3 = 200

R4 = 100

R4 = 200

R4 = 200

R3 = 100

(b)

(c)

(d)

4 (a)

Figure 8-22. (a) Two CPUs writing and two CPUs reading a common memory word. (b) – (d) Three possible ways the two writes and four reads might be interleaved in time.

However—and this is the essence of sequential consistency—no matter what, a sequentially consistent memory will never allow CPU 3 to get (100, 200) while CPU 4 gets (200, 100). If this were to occur, it would mean that according to CPU 3, the write of 100 by CPU 1 completed after the write of 200 by CPU 2. That is fine. But it would also mean that according to CPU 4, the write of 200 by CPU 2 completed before the write of 100 by CPU 1. By itself, this result is also possible. The problem is that sequential consistency guarantees that there is a single global ordering of all writes that is visible to all CPUs. If CPU 3 observes that 100 was written first, then CPU 4 must also see this order. While sequential consistency is not as powerful a rule as strict consistency, it is still very useful. In effect, it says that when multiple events are happening concurrently, there is some true order in which they occur, possibly determined by timing and chance, but a true ordering exists and all processors observe this same order. Although this statement may seem obvious, below we will discuss consistency models that do not even guarantee this much.

592

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Processor Consistency A looser consistency model, but one that is easier to implement on large multiprocessors, is processor consistency (Goodman, 1989). It has two properties: 1. Writes by any CPU are seen by all CPUs in the order they were issued. 2. For every memory word, all CPUs see all writes to it in the same order. Both of these points are important. The first point says that if CPU 1 issues writes with values 1A, 1B, and 1C to some memory location in that sequence, then all other processors see them in that order too. In other words, any other processor in a tight loop observing 1A, 1B, and 1C by reading the words written will never see the value written by 1B and then see the value written by 1A, and so on. The second point is needed to require every memory word to have an unambiguous value after several CPUs write to it and finally stop. Everyone must agree on who went last. Even with these constraints, the designer has a lot of flexibility. Consider what happens if CPU 2 issues writes 2A, 2B, and 2C concurrently with CPU 1’s three writes. Other CPUs that are busily reading memory will observe some interleaving of the six writes, such as 1A, 1B, 2A, 2B, 1C, 2C or 2A, 1A, 2B, 2C, 1B, 1C or many others. Processor consistency does not guarantee that every CPU sees the same ordering (unlike sequential consistency, which does make this guarantee). Thus it is perfectly legitimate for the hardware to behave in such a way that some CPUs see the first ordering above, some see the second, and some see yet other ones. What is guaranteed is that no CPU will see a sequence in which 1B comes before 1A, and so on. The order each CPU does its writes is observed everywhere. It is worth noting that some authors define processor consistency differently and do not require the second condition. Weak Consistency Our next model, weak consistency, does not even guarantee that writes from a single CPU are seen in order (Dubois et al., 1986). In a weakly consistent memory, one CPU might see 1A before 1B and another CPU might see 1A after 1B. However, to add some order to the chaos, weakly consistent memories have synchronization variables or a synchronization operation. When a synchronization is executed, all pending writes are finished and no new ones are started until all the old ones are done and the synchronization itself is done. In effect, a synchronization ‘‘flushes the pipeline’’ and brings the memory to a stable state with no operations pending. Synchronization operations are themselves sequentially consistent, that is, when multiple CPUs issue them, some order is chosen, but all CPUs see the same order.

SEC. 8.3

593

SHARED-MEMORY MULTIPROCESSORS

In weak consistency, time is divided into well-defined epochs delimited by the (sequentially consistent) synchronizations, as illustrated in Fig. 8-23. No relative order is guaranteed for 1A and 1B and different CPUs may see the two writes in different order, that is one CPU may see 1A then 1B and another CPU may see 1B then 1A. This situation is permitted. However, all CPUs see 1B before 1C because the first synchronization operation forces 1A, 1B, and 2A to complete before 1C, 2B, 3A, or 3B are allowed to start. Thus by doing synchronization operations, software can force some order on the sequence of events, although not at zero cost since flushing the memory pipeline does take some time. Write CPU A CPU B

1A

1B

2A

CPU C

1C

1D 1E

2B

2C

3A

3B

1F 2D

3C

Synchronization point Time

Figure 8-23. Weakly consistent memory uses synchronization operations to divide time into sequential epochs.

Release Consistency Weak consistency has the problem that it is quite inefficient because it must finish off all pending memory operations and hold all new ones until the current ones are done. Release consistency improves matters by adopting a model akin to critical sections (Gharachorloo et al., 1990). The idea behind this model is that when a process exits a critical region it is not necessary to force all the writes to complete immediately. It is only necessary to make sure that they are done before any process enters that critical region again. In this model, the synchronization operation offered by weak consistency is split into two different operations. To read or write a shared data variable, a CPU (i.e., its software) must first do an acquire operation on the synchronization variable to get exclusive access to the shared data. Then the CPU can use them as it wishes, reading and writing them at will. When it is done, the CPU does a release operation on the synchronization variable to indicate that it is finished. The release does not force pending writes to complete, but it itself does not complete until all previously issued writes are done. Furthermore, new memory operations are not prevented from starting immediately.

594

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

When the next acquire is issued, a check is made to see if all previous release operations have completed. If not, the acquire is held up until they are all done (and hence all the writes done before them are all completed). In this way, if the next acquire occurs sufficiently long after the most recent release, it does not have to wait before starting and the critical region can be entered without delay. If it occurs too soon after a release, the acquire (and all the instructions following it) will be delayed until all pending releases are completed, thus guaranteeing that the variables in the critical section have been updated. This scheme is slightly more complicated than weak consistency, but it has the significant advantage of not delaying instructions as often in order to maintain consistency. Memory consistency is not a done deal. Researchers are still proposing new models (Cain and Lipasti, 2004; and Hammond et al., 2004).

8.3.3 UMA Symmetric Multiprocessor Architectures The simplest multiprocessors are based on a single bus, as illustrated in Fig. 8-24(a). Two or more CPUs and one or more memory modules all use the same bus for communication. When a CPU wants to read a memory word, it first checks to see if the bus is busy. If the bus is idle, the CPU puts the address of the word it wants on the bus, asserts a few control signals, and waits until the memory puts the desired word on the bus.

CPU

CPU

M

Shared memory

Private memory

Shared memory CPU

CPU

M

CPU

CPU

M

Cache

Bus (a)

(b)

(c)

Figure 8-24. Three bus-based multiprocessors. (a) Without caching. (b) With caching. (c) With caching and private memories.

If the bus is busy when a CPU wants to read or write memory, the CPU just waits until the bus becomes idle. Herein lies the problem with this design. With two or three CPUs, contention for the bus will be manageable; with 32 or 64 it will be unbearable. The system will be totally limited by the bandwidth of the bus, and most of the CPUs will be idle most of the time. The solution to this problem is to add a cache to each CPU, as depicted in Fig. 8-24(b). The cache can be inside the CPU chip, next to the CPU chip, on the processor board, or some combination of all three. Since many reads can now be satisfied out of the local cache, there will be much less bus traffic, and the system can support more CPUs. Thus caching is a big win here.

SEC. 8.3

SHARED-MEMORY MULTIPROCESSORS

595

Yet another possibility is the design of Fig. 8-24(c), in which each CPU has not only a cache, but also a local, private memory which it accesses over a dedicated (private) bus. To use this configuration optimally, the compiler should place all the program text, strings, constants and other read-only data, stacks, and local variables in the private memories. The shared memory is then only used for writable shared variables. In most cases, this careful placement will greatly reduce bus traffic, but it does require active cooperation from the compiler. Snooping Caches While the performance arguments given above are certainly true, we have glossed over a fundamental problem a bit too quickly. Suppose that memory is sequentially consistent. What happens if CPU 1 has a line in its cache, and then CPU 2 tries to read a word in the same cache line? In the absence of any special rules, it, too, would get a copy in its cache. In principle, having the same line cached twice is acceptable. Now suppose that CPU 1 modifies the line and then immediately thereafter, CPU 2 reads its copy of the line from its cache. It will get stale data, thus violating the contract between the software and memory. The program running on CPU 2 will not be happy. This problem, known in the literature as the cache coherence or cache consistency problem, is extremely serious. Without a solution, caching cannot be used, and bus-oriented multiprocessors would be limited to two or three CPUs. As a consequence of its importance, many solutions have been proposed over the years (e.g., Goodman, 1983; and Papamarcos and Patel, 1984). Although all these caching algorithms, called cache coherence protocols, differ in the details, all of them prevent different versions of the same cache line from appearing simultaneously in two or more caches. In all solutions, the cache controller is specially designed to allow it to eavesdrop on the bus, monitoring all bus requests from other CPUs and caches and taking action in certain cases. These devices are called snooping caches or sometimes snoopy caches because they ‘‘snoop’’ on the bus. The set of rules implemented by the caches, CPUs, and memory for preventing different versions of the data from appearing in multiple caches forms the cache coherence protocol. The unit of transfer and storage for a cache is called a cache line and is typically 32 or 64 bytes. The simplest cache coherence protocol is called write through. It can best be understood by distinguishing the four cases shown in Fig. 8-25. When a CPU tries to read a word that is not in its cache (i.e., a read miss), its cache controller loads the line containing that word into the cache. The line is supplied by the memory, which in this protocol is always up to date. Subsequent reads (i.e., read hits) can be satisfied out of the cache. On a write miss, the word that has been modified is written to main memory. The line containing the word referenced is not loaded into the cache. On a write

596

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

############################################################## ! Action ! ! ! Local request Remote request ############################################################## ! ! ! ! ! Read miss ! Fetch data from memory ! ! ############################################################## ! Read hit ! Use data from local cache ! ! ############################################################## ! ! ! ! !############################################################## ! ! Write miss ! Update data in memory ! ! ! ! Write hit Update cache and memory Invalidate cache entry !############################################################## ! ! ! Figure 8-25. The write-through cache coherence protocol. The empty boxes indicate that no action is taken.

hit, the cache is updated and the word is written through to main memory in addition. The essence of this protocol is that all write operations result in the word being written going through to memory to keep memory up to date at all times. Now let us look at all these actions again, but this time from the snooper’s point of view, shown in the right-hand column of Fig. 8-25. Let us call the cache performing the actions cache 1 and the snooping cache cache 2. When cache 1 misses on a read, it makes a bus request to fetch a line from memory. Cache 2 sees this but does nothing. When cache 1 has a read hit, the request is satisfied locally, and no bus request occurs, so cache 2 is not aware of cache 1’s read hits. Writes are more interesting. If CPU 1 does a write, cache 1 will make a write request on the bus, both on misses and on hits. On all writes, cache 2 checks to see if it has the word being written. If not, from its point of view this is a remote request/write miss and it does nothing. (To clarify a subtle point, note that in Fig. 8-25 a remote miss means that the word is not present in the snooper’s cache; it does not matter whether it was in the originator’s cache or not. Thus a single request may be a hit locally and a miss at the snooper, or vice versa.) Now suppose that cache 1 writes a word that is present in cache 2’s cache (remote request/write hit). If cache 2 does nothing, it will have stale data, so it marks the cache entry containing the newly modified word as being invalid. In effect, it removes the item from the cache. Because all caches snoop on all bus requests, whenever a word is written, the net effect is to update it in the originator’s cache, update it in memory, and purge it from all the other caches. In this way, inconsistent versions are prevented. Of course, cache 2’s CPU is free to read the same word on the very next cycle. In that case, cache 2 will read the word from memory, which is up to date. At that point, cache 1, cache 2, and the memory will all have identical copies of it. If either CPU does a write now, the other one’s cache will be purged, and memory will be updated. Many variations on this basic protocol are possible. For example, on a write hit, the snooping cache normally invalidates its entry containing the word being written. Alternatively, it could accept the new value and update its cache instead of marking it as invalid. Conceptually, updating the cache is the same as invalidating it followed by reading the word from memory. In all cache protocols, a

SEC. 8.3

SHARED-MEMORY MULTIPROCESSORS

597

choice must be made between an update strategy and an invalidate strategy. These protocols perform differently under different loads. Update messages carry payloads and are thus larger than invalidates but may prevent future cache misses. Another variant is loading the snooping cache on write misses. The correctness of the algorithm is not affected by loading it, only the performance. The question is: ‘‘What is the probability that a word just written will be written again soon?’’ If it is high, there is something to be said for loading the cache on write misses, known as a write-allocate policy. If it is low, it is better not to update on write misses. If the word is read soon, it will be loaded by the read miss anyway; little is gained by loading it on the write miss. As with many simple solutions, this one is inefficient. Every write operation goes to memory over the bus, so with a modest number of CPUs, the bus will still become a bottleneck. To keep the bus traffic within bounds, other cache protocols have been devised. They all have the property that not all writes go directly through to memory. Instead, when a cache line is modified, a bit is set inside the cache noting that the cache line is correct but memory is not. Eventually, such a dirty line has to be written back to memory, but possibly after many writes have been made to it. This type of protocol is known as a write-back protocol. The MESI Cache Coherence Protocol One popular write-back cache coherence protocol is called MESI, after the initials of the names of the four states (M, E, S, and I) that it uses (Papamarcos and Patel, 1984). It is based on the earlier write-once protocol (Goodman, 1983). The MESI protocol is used by the Pentium 4 and many other CPUs for snooping on the bus. Each cache entry can be in one of the following four states: 1. Invalid 2. Shared

– The cache entry does not contain valid data. – Multiple caches may hold the line; memory is up to date.

3. Exclusive – No other cache holds the line; memory is up to date. 4. Modified – The entry is valid; memory is invalid; no copies exist. When the CPU is initially booted, all cache entries are marked invalid. The first time memory is read, the line referenced is fetched into the cache of the CPU reading memory and marked as being in the E (exclusive) state, since it is the only copy in a cache, as illustrated in Fig. 8-26(a) for the case of CPU 1 reading line A. Subsequent reads by that CPU use the cached entry and do not go over the bus. Another CPU may also fetch the same line and cache it, but by snooping, the original holder (CPU 1) sees that it is no longer alone and announces on the bus that it also has a copy. Both copies are marked as being in the S (shared) state, as shown in Fig. 8-26(b). In other words, the S state means that the line is in one or

598

(a)

PARALLEL COMPUTER ARCHITECTURES

CPU 1

CPU 2

CPU 3

Memory

CHAP. 8

CPU 1 reads block A

A Exclusive Bus

Cache

(b)

CPU 1

CPU 2

CPU 3

Memory

CPU 2 reads block A

A Shared

Shared Bus

(c)

CPU 1

CPU 2

CPU 3

Memory

CPU 2 writes block A

A Modified Bus

(d)

CPU 1

CPU 2

CPU 3

A

A Shared

Memory

CPU 3 reads block A

Shared Bus

(e)

CPU 1

CPU 2

CPU 3

Memory

CPU 2 writes block A

A Modified Bus

(f)

CPU 1

CPU 2

CPU 3

Memory

CPU 1 writes block A

A Modified Bus

Figure 8-26. The MESI cache coherence protocol.

SEC. 8.3

SHARED-MEMORY MULTIPROCESSORS

599

more caches for reading and memory is up to date. Subsequent reads by a CPU to a line it has cached in the S state do not use the bus and do not cause the state to change. Now consider what happens if CPU 2 writes to the cache line it is holding in S state. It puts out an invalidate signal on the bus, telling all other CPUs to discard their copies. The copy cached now goes to M (modified) state, as shown in Fig. 8-26(c). The line is not written to memory. It is worth noting that if a line is in E state when it is written, no bus signal is needed to invalidate other caches because it is known that no other copies exist. Next consider what happens if CPU 3 reads the line. CPU 2, which now owns the line, knows that the copy in memory is not valid, so it asserts a signal on the bus telling CPU 3 to please wait while it writes its line back to memory. When it is finished, CPU 3 fetches a copy, and the line is marked as shared in both caches, as shown in Fig. 8-26(d). After that, CPU 2 writes the line again, which invalidates the copy in CPU 3’s cache, as shown in Fig. 8-26(e). Finally, CPU 1 writes to a word in the line. CPU 2 sees that a write is being attempted and asserts a bus signal telling CPU 1 to please wait while it writes its line back to memory. When it is finished, it marks its own copy as invalid, since it knows another CPU is about to modify it. At this point we have the situation in which a CPU is writing to an uncached line. If the write-allocate policy is in use, the line will be loaded into the cache and marked as being in the M state, as shown in Fig. 8-26(f). If the write-allocate policy is not in use, the write will go directly to memory and the line will not be cached anywhere. UMA Multiprocessors Using Crossbar Switches Even with all possible optimizations, the use of a single bus limits the size of a UMA multiprocessor to about 16 or 32 CPUs. To go beyond that, a different kind of interconnection network is needed. The simplest circuit for connecting n CPUs to k memories is the crossbar switch, shown in Fig. 8-27. Crossbar switches have been used for decades within telephone switching exchanges to connect a group of incoming lines to a set of outgoing lines in an arbitrary way. At each intersection of a horizontal (incoming) and vertical (outgoing) line is a crosspoint. A crosspoint is a small switch that can be electrically opened or closed, depending on whether the horizontal and vertical lines are to be connected or not. In Fig. 8-27(a) we see three crosspoints closed simultaneously, allowing connections between the (CPU, memory) pairs (001, 000), (101, 101), and (110, 010) at the same time. Many other combinations are also possible. In fact, the number of combinations is equal to the number of different ways eight rooks can be safely placed on a chess board. One of the nicest properties of the crossbar switch is that it is a nonblocking network, meaning that no CPU is ever denied the connection it needs because

600

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Crosspoint switch is open

111

110

101

100

011

010

001

000

Memories

000 001

CPUs

010

(b)

011

Crosspoint switch is closed

100 101 110 111 Closed crosspoint switch

Open crosspoint switch

(c)

(a)

Figure 8-27. (a) An 8 × 8 crossbar switch. (b) An open crosspoint. (c) A closed crosspoint.

some crosspoint or line is already occupied (assuming the memory module itself is available). Furthermore, no advance planning is needed. Even if seven arbitrary connections are already set up, it is always possible to connect the remaining CPU to the remaining memory. We will later see interconnection schemes that do not have these properties. One of the worst properties of the crossbar switch is the fact that the number of crosspoints grows as n 2 . For medium-sized systems, a crossbar design is workable. We will discuss one such design, the Sun Fire E25K, later in this chapter. However, with 1000 CPUs and 1000 memory modules, we need a million crosspoints. Such a large crossbar switch is not feasible. We need something quite different. UMA Multiprocessors Using Multistage Switching Networks That ‘‘something quite different’’ can be based on the humble 2 × 2 switch shown in Fig. 8-28(a). This switch has two inputs and two outputs. Messages arriving on either input line can be switched to either output line. For our purposes here, messages will contain up to four parts, as shown in Fig. 8-28(b). The

SEC. 8.3

601

SHARED-MEMORY MULTIPROCESSORS

Module field tells which memory to use. The Address specifies an address within a module. The Opcode gives the operation, such as READ or WRITE. Finally, the optional Value field may contain an operand, such as a 32-bit word to be written on a WRITE. The switch inspects the Module field and uses it to determine if the message should be sent on X or on Y. A

X

B

Y

Module

Address

(a)

Opcode

Value

(b)

Figure 8-28. (a) A 2 × 2 switch. (b) A message format.

Our 2 × 2 switches can be arranged in many ways to build larger multistage switching networks One possibility is the no-frills, economy class omega network, illustrated in Fig. 8-29. Here we have connected eight CPUs to eight memories using 12 switches. More generally, for n CPUs and n memories we would need log2 n stages, with n/2 switches per stage, for a total of (n/2)log2 n switches, which is a lot better than n 2 crosspoints, especially for large values of n. 3 Stages Memories

CPUs 000 001

1A

b

010

1B

011

101

111

b

000

3A b

2B

3B

1C

2C

3C

1D

2D

3D

a

011 100 101 a

a

a

001 010

b

100

110

2A

110 111

Figure 8-29. An omega switching network.

The wiring pattern of the omega network is often called the perfect shuffle, since the mixing of the signals at each stage resembles a deck of cards being cut in half and then mixed card-for-card. To see how the omega network works, suppose that CPU 011 wants to read a word from memory module 110. The CPU sends a READ message to switch 1D containing 110 in the Module field. The switch takes the first (i.e., leftmost) bit of 110 and uses it for routing. A 0 routes to the upper output and a 1 routes to the lower one. Since this bit is a 1, the message is routed via the lower output to 2D.

602

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

All the second-stage switches, including 2D, use the second bit for routing. This, too, is a 1, so the message is now forwarded via the lower output to 3D. Here the third bit is tested and found to be a 0. Consequently, the message goes out on the upper output and arrives at memory 110, as desired. The path followed by this message is marked in Fig. 8-29 by the letter a. As the message moves through the switching network, the bits at the left-hand end of the module number are no longer needed. They can be put to good use by recording the incoming line number there, so the reply can find its way back. For path a, the incoming lines are 0 (upper input to 1D), 1 (lower input to 2D), and 1 (lower input to 3D), respectively. The reply is routed back using 011, only reading it from right to left this time. At the same time all this is going on, CPU 001 wants to write a word to memory module 001. An analogous process happens here, with the message routed via the upper, upper, and lower outputs, respectively, marked by the letter b. When it arrives, its Module field reads 001, representing the path it took. Since these two requests do not use any of the same switches, lines, or memory modules, they can proceed in parallel. Now consider what would happen if CPU 000 simultaneously wanted to access memory module 000. Its request would come into conflict with CPU 001’s request at switch 3A. One of them would have to wait. Unlike the crossbar switch, the omega network is a blocking network. Not every set of requests can be processed simultaneously. Conflicts can occur over the use of a wire or a switch, as well as between requests to memory and replies from memory. It is clearly desirable to spread the memory references uniformly across the modules. One common technique is to use the low-order bits as the module number. Consider, for example, a byte-oriented address space for a computer that mostly accesses 32-bit words. The 2 low-order bits will usually be 00, but the next 3 bits will be uniformly distributed. By using these 3 bits as the module number, consecutively addressed words will be in consecutive modules. A memory system in which consecutive words are in different modules is said to be interleaved. Interleaved memories maximize parallelism because most memory references are to consecutive addresses. It is also possible to design switching networks that are nonblocking and which offer multiple paths from each CPU to each memory module, to spread the traffic better.

8.3.4 NUMA Multiprocessors It should be clear by now that single-bus UMA multiprocessors are generally limited to no more than a few dozen CPUs and crossbar or switched multiprocessors need a lot of (expensive) hardware and are not that much bigger. To get to more than 100 CPUs, something has to give. Usually, what gives is the idea that all memory modules have the same access time. This concession leads to the idea of NUMA (NonUniform Memory Access) multiprocessors. Like their UMA

SEC. 8.3

603

SHARED-MEMORY MULTIPROCESSORS

cousins, they provide a single address space across all the CPUs, but unlike the UMA machines, access to local memory modules is faster than access to remote ones. Thus all UMA programs will run without change on NUMA machines, but the performance will be worse than on a UMA machine at the same clock speed. NUMA machines have three key characteristics that all of them possess and which together distinguish them from other multiprocessors: 1. There is a single address space visible to all CPUs. 2. Access to remote memory is done using LOAD and STORE instructions. 3. Access to remote memory is slower than access to local memory. When the access time to remote memory is not hidden (because there is no caching), the system is called NC-NUMA. When coherent caches are present, the system is called CC-NUMA (at least by the hardware people). The software people often call it hardware DSM because it is basically the same as software distributed shared memory but implemented by the hardware using a small page size. One of the first NC-NUMA machines (although the name had not yet been coined) was the Carnegie-Mellon Cm*, illustrated in simplified form in Fig. 8-30 (Swan et al., 1977). It consisted of a collection of LSI-11 CPUs, each with some memory addressed over a local bus. (The LSI-11 was a single-chip version of the DEC PDP-11, a minicomputer popular in the 1970s). In addition, the LSI-11 systems were connected by a system bus. When a memory request came into the (specially modified) MMU, a check was made to see if the word needed was in the local memory. If so, a request was sent over the local bus to get the word. If not, the request was routed over the system bus to the system containing the word, which then responded. Of course, the latter took much longer than the former. While a program could run happily out of remote memory, it took 10 times longer to execute than the same program running out of local memory. CPU Memory

MMU

Local bus

CPU Memory

Local bus

CPU Memory

Local bus

CPU Memory

Local bus

System bus

Figure 8-30. A NUMA machine based on two levels of buses. The Cm* was the first multiprocessor to use this design.

Memory coherence is guaranteed in an NC-NUMA machine because there is no caching present. Each word of memory lives in exactly one location, so there

604

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

is no danger of one copy having stale data: there are no copies. Of course, it now matters a great deal which page is in which memory because the performance penalty for being in the wrong place is so high. Consequently, NC-NUMA machines use elaborate software to move pages around to maximize performance. Typically, there is a daemon process called a page scanner that runs every few seconds. Its job is to examine the usage statistics and move pages around in an attempt to improve performance. If a page appears to be in the wrong place, the page scanner unmaps it so that the next reference to it will cause a page fault. When the fault occurs, a decision is made about where to place the page, possibly in a different memory from where it was before. To prevent thrashing, usually there is some rule saying that once a page is placed, it is frozen in place for a time ∆T. Various algorithms have been studied, but the conclusion is that no one algorithm performs best under all circumstances (LaRowe and Ellis, 1991). Cache Coherent NUMA Multiprocessors Multiprocessor designs such as that of Fig. 8-30 do not scale well because they do not do caching. Having to go to the remote memory every time a nonlocal memory word is accessed is a major performance hit. However, if caching is added, then cache coherence must also be added. One way to provide cache coherence is to snoop on the system bus. Technically, doing this is not difficult, but beyond a certain number of CPUs, it becomes infeasible. To build really large multiprocessors, a fundamentally different approach is needed. The most popular approach for building large CC-NUMA (Cache Coherent NUMA) multiprocessors currently is the directory-based multiprocessor. The idea is to maintain a database telling where each cache line is and what its status is. When a cache line is referenced, the database is queried to find out where it is and whether it is clean or dirty (modified). Since this database must be queried on every single instruction that references memory, it must be kept in extremely-fast special-purpose hardware that can respond in a fraction of a bus cycle. To make the idea of a directory-based multiprocessor somewhat more concrete, let us consider as a simple (hypothetical) example, a 256-node system, each node consisting of one CPU and 16 MB of RAM connected to the CPU via a local bus. The total memory is 232 bytes, divided up into 226 cache lines of 64 bytes each. The memory is statically allocated among the nodes, with 0–16M in node 0, 16M–32M in node 1, and so on. The nodes are connected by an interconnection network, as shown in Fig. 8-31(a). The interconnection network could be a grid, hypercube, or other topology. Each node also holds the directory entries for the 218 64-byte cache lines comprising its 224 byte memory. For the moment, we will assume that a line can be held in at most one cache. To see how the directory works, let us trace a LOAD instruction from CPU 20 that references a cached line. First the CPU issuing the instruction presents it to its MMU, which translates it to a physical address, say, 0x24000108. The MMU

SEC. 8.3

605

SHARED-MEMORY MULTIPROCESSORS Node 0

Node 1

CPU Memory

CPU Memory

Local bus

Local bus

Node 255 CPU Memory

Directory

…

Local bus

Interconnection network (a) Bits

8

18

6

Node

Block

Offset

(b)

218-1

4 3 2 1 0

0 0 1 0 0

82

(c)

Figure 8-31. (a) A 256-node directory-based multiprocessor. (b) Division of a 32-bit memory address into fields. (c) The directory at node 36.

splits this address into the three parts shown in Fig. 8-31(b). In decimal, the three parts are node 36, line 4, and offset 8. The MMU sees that the memory word referenced is from node 36, not node 20, so it sends a request message through the interconnection network to the line’s home node, 36, asking whether its line 4 is cached, and if so, where. When the request arrives at node 36 over the interconnection network, it is routed to the directory hardware. The hardware indexes into its table of 218 entries, one for each of its cache lines and extracts entry 4. From Fig. 8-31(c) we see that the line is not cached, so the hardware fetches line 4 from the local RAM, sends it back to node 20, and updates directory entry 4 to indicate that the line is now cached at node 20. Now let us consider a second request, this time asking about node 36’s line 2. From Fig. 8-31(c) we see that this line is cached at node 82. At this point the hardware could update directory entry 2 to say that the line is now at node 20 and then send a message to node 82 instructing it to pass the line to node 20 and invalidate its cache. Note that even a so-called ‘‘shared-memory multiprocessor’’ has a lot of message passing going on under the hood. As a quick aside, let us calculate how much memory is being taken up by the directories. Each node has 16 MB of RAM and 218 9-bit entries to keep track of

606

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

that RAM. Thus the directory overhead is about 9 × 218 bits divided by 16 MB or about 1.76 percent, which is generally acceptable (although it has to be highspeed memory, which increases its cost). Even with 32-byte cache lines the overhead would only be 4 percent. With 128-byte cache lines, it would be under 1 percent. An obvious limitation of this design is that a line can be cached at only one node. To allow lines to be cached at multiple nodes, we would need some way of locating all of them, for example, to invalidate or update them on a write. Various options are possible to allow caching at several nodes at the same time. One possibility is to give each directory entry k fields for specifying other nodes, thus allowing each line to be cached at up to k nodes. A second possibility is to replace the node number in our simple design with a bit map, with one bit per node. In this option there is no limit on how many copies there can be, but there is a substantial increase in overhead. Having a directory with 256 bits for each 64-byte (512-bit) cache line implies an overhead of over 50 percent. A third possibility is to keep one 8-bit field in each directory entry and use it as the head of a linked list that threads all the copies of the cache line together. This strategy requires extra storage at each node for the linked list pointers, and it also requires following a linked list to find all the copies when that is needed. Each possibility has its own advantages and disadvantages, and all three of them have been used in real systems. Another improvement to the directory design is to keep track of whether the cache line is clean (home memory is up to date) or dirty (home memory is not up to date). If a read request comes in for a clean cache line, the home node can satisfy the request from memory, without having to forward it to a cache. A read request for a dirty cache line, however, must be forwarded to the node holding the cache line because only it has a valid copy. If only one cache copy is allowed, as in Fig. 8-31, there is no real advantage to keeping track of its cleanliness, because any new request requires a message to be sent to the existing copy to invalidate it. Of course, keeping track of whether each cache line is clean or dirty implies that when a cache line is modified, the home node has to be informed, even if only one cache copy exists. If multiple copies exist, modifying one of them requires the rest to be invalidated, so some protocol is needed to avoid race conditions. For example, to modify a shared cache line, one of the holders might have to request exclusive access before modifying it. Such a request would cause all other copies to be invalidated before permission was granted. Other performance optimizations for CC-NUMA machines are discussed in (Stenstrom et al., 1997). The Sun Fire E25K NUMA Multiprocessor As an example of a shared-memory NUMA multiprocessor, let us study the Sun Microsystems Sun Fire family. Although this family contains various models, we will focus on the E25K, which has 72 UltraSPARC IV CPU chips. An

SEC. 8.3

607

SHARED-MEMORY MULTIPROCESSORS

UltraSPARC IV is essentially a pair of UltraSPARC III Cu processors that share a common cache and memory. The E15K is essentially the same system except with uniprocessor instead of dual processor CPU chips. Smaller members exist as well, but from our point of view, what is interesting is how the one with the most CPUs works. The E25K system consists of up to 18 boardsets, each boardset consisting of a CPU-memory board, an I/O board with four PCI slots, and an expander board that couples the CPU-memory board with the I/O board and joins the pair to the centerplane, which holds the boards and contains the switching logic. Each CPUmemory board contains four CPU chips and four 8-GB RAM modules. Consequently, each CPU-memory board on the E25K holds eight CPUs and 32 GB of RAM (four CPUs and four 32 GB of RAM on the E15K). A full E25K thus contains 144 CPUs, 576 GB of RAM, and 72 PCI slots. It is illustrated in Fig. 8-32. Interestingly enough, the number 18 was chosen due to packaging constraints: a system with 18 boardsets was the largest one that could fit through a doorway in one piece. While programmers just think about 0s and 1s, engineers have to worry about things like how the customer will get the product through the door and into the building. System control board

er

nt

Ce

9 CPU-memory boards

ne

pla

PCI cont.

18 x 18 data crossbar

PCI cont.

Snooping bus

CPU

Expander board

Memory

Directory Control

CPU

18 x 18 resp crossbar

Memory

18 x 18 addr crossbar

CPU

Expander board

Memory

Directory Control

CPU Snooping bus

Memory

CPU

Memory

CPU

Memory

CPU

Memory

CPU

Memory

PCI cont. PCI cont.

9 I/O boards

PCI slot

Figure 8-32. The Sun Microsystems E25K multiprocessor.

The centerplane is composed of a set of three 18 × 18 crossbar switches for connecting the 18 boardsets. One crossbar is for the address lines, one is for

608

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

responses, and one is for data transfer. In addition to the 18 expander boards, the centerplane also has a system control boardset plugged into it. This boardset has a single CPU, but also interfaces to the CD-ROM, tape, serial lines, and other peripheral devices needed for booting, maintaining, and controlling the system. The heart of any multiprocessor is the memory subsystem. How does one connect 144 CPUs to the distributed memory? The straightforward ways—a big shared snooping bus or a 144 × 72 crossbar switch—do not work well. The former fails due to the bus being a bottleneck and the latter fails because the switch is too difficult and expensive to build. Thus large multiprocessors such as the E25K are forced to use a more complex memory subsystem. At the boardset level, snooping logic is used so all local CPUs can check all memory requests coming from the boardset to references to blocks they currently have cached. Thus when a CPU needs a word from memory, it first converts the virtual address to a physical address and checks its cache. (Physical addresses are 43 bits, but packaging restrictions limit memory to 576 GB.) If the cache block it needs is in its own cache, the word is returned. Otherwise, the snooping logic checks if a copy of that word is available somewhere else on the boardset. If so, the request is satisfied. If not, the request is passed on via the 18 × 18 address crossbar switch as described below. The snooping logic can do one snoop per clock cycle. The system clock runs at 150 MHz, so it is possible to perform 150 million snoops/sec per boardset or 2.7 billion snoops/sec system wide. Although the snooping logic is logically a bus, as portrayed in Fig. 8-32, physically, it is a device tree, with commands being relayed up and down the tree. When a CPU or PCI board puts out an address, it goes to an address repeater via a point-to-point connection, as shown in Fig. 8-33. The two address repeaters converge on the expander board, where the addresses are sent back down the tree for each device to check for hits. This arrangement is used to avoid having a bus that involves three boards. Data transfers use a four-level interconnect as depicted in Fig. 8-33. This design was chosen for high performance. At level 0, pairs of CPU chips and memories are connected by a small crossbar switch that also has a connection to level 1. The two groups of CPU-memory pairs are connected by a second crossbar switch at level 1. The crossbar switches are custom ASICs. For all of them, all the inputs are available on both the rows and the columns, although not all combinations are used (or even make sense). All the switching logic on the boards is built from 3 × 3 crossbars. Each boardset consists of three boards: the CPU-memory board, the I/O board, and the expander board, which connects the other two. The level 2 interconnect is another 3 × 3 crossbar switch (on the expander board) that joins the actual memory to the I/O ports (which are memory mapped on all UltraSPARCs). All data transfers to or from the boardset, whether to memory or to an I/O port, pass through the level 2 switch. Finally, data that has to be transferred to or from a remote board passes through an 18 × 18 data crossbar switch at level 3. Data

SEC. 8.3

3x3 Data switch

5x5 Data switch

M

C 5x5 Data switch

Memory M

C

M

Addr rep.

PCI cont.

3x3 Data switch PCI cont.

Level 1 I/O board

C

CPU-memory board

Snooping domain

3 x 3 Data switch

C

Level 2

18 x 18 Dat crossbar switch

Directory and snooping management

Addr rep.

Level 3

Expander board

18 x 18 Response crossbar switch

Centerplane

Directory domain

18 x 18 Address crossbar switch

CPU

609

SHARED-MEMORY MULTIPROCESSORS

Level 0

M

Figure 8-33. The Sun Fire E25K uses a four-level interconnect. Dashed lines are address paths. Solid lines are data paths.

transfers are done 32 bytes at a time, so it takes two clock cycles to transfer 64 bytes, the usual transfer unit. Having looked at how the components are arranged, let us now turn our attention to how the shared memory operates. At the bottom level, the 576 GB of memory is split into 229 blocks of 64 bytes each. These blocks are the atomic units of the memory system. Each block has a home board where it lives when not in use elsewhere. Most blocks are on their home board most of the time. However, when a CPU needs a memory block, either from its own board or one of the 17 remote ones, it first requests a copy for its own cache, then accesses the cached copy. Although each CPU chip on the E25K contains two CPUs, they share a single physical cache and thus share all the blocks contained in it. Each memory block and cache line of each CPU chip can be in one of three states: 1. Exclusive access (for writing). 2. Shared access (for reading). 3. Invalid (i.e., empty). When a CPU needs to read or write a memory word, it first checks its own cache. Failing to find the word there, it issues a local request for the physical address that is broadcast only on its own boardset. If a cache on the boardset has the needed

610

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

line, the snooping logic detects the hit and responds to the request. If the line is in exclusive mode, it is transferred to the requestor and the original copy marked invalid. If it is in shared mode, the cache does not respond since memory always responds when a cache line is clean. If the snooping logic cannot find the cache line or it is present and shared, it sends a request over the centerplane to the home board asking where the memory block is. The state of each memory block is stored in the block’s ECC bits, so the home board can immediately determine its state. If the block is either unshared or shared with one or more remote boards, the home memory will be up to date, and the request can be satisfied from the home board’s memory. In this case, a copy of the case line is transmitted over the data crossbar switch in two clock cycles, eventually arriving at the requesting CPU. If the request was for reading, an entry is made in the directory at the home board noting that a new customer is sharing the cache line and the transaction is finished. However, if the request was for writing, an invalidation message has to be sent to all other boards (if any) holding a copy of it. In this way, the board making the write request ends up with the only copy. Now consider the case in which the requested block is in exclusive state located on a different board. When the home board gets the request, it looks up the location of the remote board in the directory and sends the requester a message telling where the cache line is. The requester now sends the request to the correct boardset. When the request arrives, the board sends back the cache line. If it was a read request, the line is marked shared and a copy sent back to the home board. If it was a write request, the responder invalidates its copy so the new requester has an exclusive copy. Since each board has 229 memory blocks, it would take a directory with 229 entries to keep track of them all in the worst case. Since the directory is much smaller than 229 , it could happen that there is no room in the directory (which is searched associatively) for some entries. In this case, the home directory has to locate the block by broadcasting a request for it to all the other 17 boards. The response crossbar switch plays a roll in the directory coherence and update protocol by handling much of the reverse traffic back to the sender. By splitting the protocol traffic over two buses (address and response) and the data over a third bus, the throughput of the system is increased. By distributing the load over multiple devices on different boards, the Sun Fire E25K is able to achieve very high performance. In addition to the 2.7 billion snoops/sec mentioned above, the centerplane can handle up to nine simultaneous transfers, with nine boards sending and nine boards receiving. Since the data crossbar is 32 bytes wide, on every clock cycle, 288 bytes can be moved through the centerplane. At a clock rate of 150 MHz, this gives a peak aggregate bandwidth of 40 GB/sec when all accesses are remote. If the software can place pages in such a way to ensure that most accesses are local, then the system bandwidth can be appreciably higher than 40 GB/sec.

SEC. 8.3

SHARED-MEMORY MULTIPROCESSORS

611

For more technical information about the Sun Fire, see (Charlesworth, 2002; and Charlesworth, 2001);

8.3.5 COMA Multiprocessors NUMA and CC-NUMA machines have the disadvantage that references to remote memory are much slower than references to local memory. In CC-NUMA, this performance difference is hidden to some extent by the caching. Nevertheless, if the amount of remote data needed greatly exceeds the cache capacity, cache misses will occur constantly and performance will be poor. Thus we have a situation that UMA machines have excellent performance but are limited in size and are quite expensive. NC-NUMA machines, scale to somewhat larger sizes but require manual or semi-automated placement of pages, often with mixed results. The problem is that it is hard to predict which pages will be needed where, and in any case, pages are often too large a unit to move around. CC-NUMA machines, such as the Sun Fire E25K, may experience poor performance if many CPUs need a lot of remote data. All in all, each of these designs has serious limitations. An alternative kind of multiprocessor tries to get around all these problems by using each CPU’s main memory as a cache. In this design, called COMA (Cache Only Memory Access), pages do not have fixed home machines, as they do in NUMA and CC-NUMA machines. In fact, pages are not significant at all. Instead, the physical address space is split into cache lines, which migrate around the system on demand. Blocks do not have home machines. Like nomads in some Third World countries, home is where you are right now. A memory that just attracts lines as needed is called an attraction memory. Using the main RAM as a big cache greatly increases the hit rate, hence the performance. Unfortunately, as usual, there is no such thing as a free lunch. COMA systems introduce two new problems: 1. How are cache lines located? 2. When a line is purged from memory, what happens if it is the last copy? The first problem relates to the fact that after the MMU has translated a virtual address to a physical address, if the line is not in the true hardware cache, there is no easy way to tell if it is in main memory at all. The paging hardware does not help here at all because each page is made up of many individual cache lines that wander around independently. Furthermore, even if it is known that a line is not in main memory, where is it then? It is not possible to just ask the home machine, because there is no home machine. Some solutions to the location problem have been proposed. To see if a cache line is in main memory, new hardware could be added to keep track of the tag for

612

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

each cached line. The MMU could then compare the tag for the line needed to the tags for all the cache lines in memory to look for a hit. This solution needs additional hardware. A somewhat different solution is to map entire pages in but not require that all the cache lines be present. In this solution, the hardware would need a bit map per page, giving one bit per cache line indicating the presence or absence of the line. In this design, called simple COMA if a cache line is present, it must be in the right position in its page, but if it is not present, any attempt to use it causes a trap to allow the software to go find it and bring it in. This leads us to finding lines that are really remote. One solution is to give each page a home machine in terms of where its directory entry is, but not where the data are. Then a message can be sent to the home machine to at least locate the cache line. Other schemes involve organizing memory as a tree and searching upward until the line is found. The second problem in the list above relates to not purging the last copy. As in CC-NUMA, a cache line may be at multiple nodes at once. When a cache miss occurs, a line must be fetched, which usually means a line must be thrown out. What happens if the line chosen happens to be the last copy? In that case, it cannot be thrown out. One solution is to go back to the directory and check to see if there are other copies. If so, the line can be safely thrown out. Otherwise, it has to be migrated somewhere else. Another solution is to label one copy of each cache line as the master copy and never throw it out. This solution avoids having to check with the directory. All in all, COMA offers promise to provide better performance than CC-NUMA, but few COMA machines have been built, so more experience is needed. The first two COMA machines built were the KSR-1 (Burkhardt et al., 1992) and the Data Diffusion Machine (Hagersten et al., 1992). A more recent machine is the SDAARC (Eschmann et al., 2002).

8.4 MESSAGE-PASSING MULTICOMPUTERS As we saw in Fig. 8-21, the two kinds of MIMD parallel processors are multiprocessors and multicomputers. In the previous section we studied multiprocessors. We saw that multiprocessors appear to the operating system as having shared memory that can be accessed using ordinary LOAD and STORE instructions. This shared memory can be implemented in many ways as we have seen, including snooping buses, data crossbars, multistage switching networks, and various directory based schemes. Nevertheless, programs written for a multiprocessor can just access any location in memory without knowing anything about the internal topology or implementation scheme. This illusion is what makes multiprocessors so attractive and why programmers like this programming model.

SEC. 8.4

613

MESSAGE-PASSING MULTICOMPUTERS

On the other hand, multiprocessors also have their limitations, which is why multicomputers are important, too. First and foremost, multiprocessors do not scale to large sizes. We saw the enormous amount of hardware Sun had to use to get the E25K to scale to 72 CPUs. In contrast, we will study a multicomputer below that has 65,536 CPUs. It will be years before anyone builds a commercial 65,536-node multiprocessor, and by then million-node multicomputers will be in use. In addition, memory contention in a multiprocessor can severely affect performance. If 100 CPUs are all trying to read and write the same variables constantly, contention for the various memories, buses, and directories can result in an enormous performance hit. As a consequence of these and other factors, there is a great deal of interest in building and using parallel computers in which each CPU has its own private memory, not directly accessible to any other CPU. They are the multicomputers. Programs on multicomputer CPUs interact using primitives like send and receive to explicitly pass messages because they cannot get at each other’s memory with LOAD and STORE instructions. This difference completely changes the programming model. Each node in a multicomputer consists of one or a few CPUs, some RAM (conceivably shared among the CPUs at that node only), a disk and/or other I/O devices, and a communication processor. The communication processors are connected by a high-speed interconnection network of the types we discussed in Sec. 8.1.2. Many different topologies, switching schemes, and routing algorithms are used. What all multicomputers have in common is that when an application program executes the send primitive, the communication processor is notified and transmits a block of user data to the destination machine (possibly after first asking for and getting permission). A generic multicomputer is shown in Fig. 8-34.

CPU

Node

Memory

… Local interconnect

Disk and I/O

…

… Local interconnect

Communication processor High-performance interconnection network

Figure 8-34. A generic multicomputer.

Disk and I/O

614

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

8.4.1 Interconnection Networks In Fig. 8-34 we saw that multicomputers are held together by interconnection networks. Now it is time to look more closely at these interconnection networks. Interestingly enough, multiprocessors and multicomputers are surprisingly similar in this respect because multiprocessors often have multiple memory modules that must also be interconnected with one another and with the CPUs. Thus the material in this section frequently applies to both kinds of systems. The fundamental reason why multiprocessor and multicomputer interconnection networks are similar is that at the very bottom both of them use message passing. Even on a single-CPU machine, when the processor wants to read or write a word, what it typically does is assert certain lines on the bus and wait for a reply. This action is fundamentally like message passing: the initiator sends a request and waits for a response. In large multiprocessors, communication between CPUs and remote memory almost always consists of the CPU sending an explicit message, called a packet, to memory requesting some data, and the memory sending back a reply packet. Topology The topology of an interconnection network describes how the links and switches are arranged, for example, as a ring or as a grid. Topological designs can be modeled as graphs, with the links as arcs and the switches as nodes, as shown in Fig. 8-35. Each node in an interconnection network (or its graph) has some number of links connected to it. Mathematicians call the number of links the degree of the node; engineers call it the fanout. In general, the greater the fanout, the more routing choices there are and the greater the fault tolerance, that is, the ability to continue functioning even if a link fails by routing around it. If every node has k arcs and the wiring is done right, it is possible to design the network so that it remains fully connected even if k − 1 links fail. Another property of an interconnection network (or its graph) is its diameter. If we measure the distance between two nodes by the number of arcs that have to be traversed to get from one to the other, then the diameter of a graph is the distance between the two nodes that are the farthest apart (i.e., have the greatest distance between them). The diameter of an interconnection network is related to the worst-case delay when sending packets from CPU to CPU or from CPU to memory because each hop across a link takes a finite amount of time. The smaller the diameter, the better the worst-case performance. Also important is the average distance between two nodes, since this relates to the average packet transit time. Yet another important property of an interconnection network is its transmission capacity, that is, how much data it can move per second. One useful measure of this capacity is the bisection bandwidth. To compute this quantity, we first have to (conceptually) partition the network into two equal (in terms of number of

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

615

Figure 8-35. Various topologies. The heavy dots represent switches. The CPUs and memories are not shown. (a) A star. (b) A complete interconnect. (c) A tree. (d) A ring. (e) A grid. (f) A double torus. (g) A cube. (h) A 4D hypercube.

nodes) but unconnected parts by removing a set of arcs from its graph. Then we compute the total bandwidth of the arcs that have been removed. There may be many different ways to partition the network into two equal parts. The bisection bandwidth is the minimum of all the possible partitions. The significance of this number is that if the bisection bandwidth is, say, 800 bits/sec, then if there is a lot

616

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

of communication between the two halves, the total throughput may be limited to only 800 bits/sec, in the worst case. Many designers believe bisection bandwidth is the most important metric of an interconnection network. Many interconnection networks are designed with the goal of maximizing the bisection bandwidth. Interconnection networks can be characterized by their dimensionality. For our purposes, the dimensionality is determined by the number of choices there are to get from the source to the destination. If there is never any choice (i.e., there is only one path from each source to each destination), the network is zero dimensional. If there is one dimension in which a choice can be made, for example, go east or go west, the network is one dimensional. If there are two axes, so a packet can go east or west, or alternatively, go north or south, the network is two dimensional, and so on. Several topologies are shown in Fig. 8-35. Only the links (lines) and switches (dots) are shown here. The memories and CPUs (not shown) would typically be attached to the switches by interfaces. In Fig. 8-35(a), we have a zero-dimensional star configuration, in which the CPUs and memories would be attached to the outer nodes, with the central one just doing switching. Although a simple design, for a large system, the central switch is likely to be a major bottleneck. Also, from a fault-tolerance perspective, this is a poor design since a single failure at the central switch completely destroys the system. In Fig. 8-35(b), we have another zero-dimensional design that is at the other end of the spectrum, a full interconnect. Here every node has a direct connection to every other node. This design maximizes the bisection bandwidth, minimizes the diameter, and is exceedingly fault tolerant (it can lose any six links and still be fully connected). Unfortunately, the number of links required for k nodes is k(k − 1)/2, which quickly gets out of hand for large k. Another topology is the tree, illustrated in Fig. 8-35(c). A problem with this design is that the bisection bandwidth is equal to the link capacity. Since there will normally be a lot of traffic near the top of the tree, the top few nodes will become bottlenecks. One way around this problem is to increase the bisection bandwidth by giving the upper links more bandwidth. For example, the lowest level links might have a capacity b, the next level might have a capacity 2b and the top-level links might each have 4b. Such a design is called a fat tree and has been used in commercial multicomputers, such as the (now-defunct) Thinking Machines’ CM-5. The ring of Fig. 8-35(d) is a one-dimensional topology by our definition because every packet sent has a choice of going left or going right. The grid or mesh of Fig. 8-35(e) is a two-dimensional design that has been used in many commercial systems. It is highly regular, easy to scale up to large sizes, and has a diameter that only increases as the square root of the number of nodes. A variant on the grid is the double torus of Fig. 8-35(f), which is a grid with the edges connected. Not only is it more fault tolerant than the grid, but the diameter is also less because the opposite corners can now communicate in only two hops.

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

617

Yet another popular topology is the three-dimensional torus. The topology here consist of a 3D-structure with nodes at the points (i, j, k) where all coordinates are integers in the range from (1, 1, 1) to (l, m, n). Each node has six neighbors, two along each axis. The nodes at the edges have links that wrap around to the opposite edge, just as with the 2D-torus. The cube of Fig. 8-35(g) is a regular three-dimensional topology. We have illustrated a 2 × 2 × 2 cube, but in the general case it could be a k × k × k cube. In Fig. 8-35(h) we have a four-dimensional cube constructed from two threedimensional cubes with the corresponding nodes connected. We could make a five-dimensional cube by cloning the structure of Fig. 8-35(h) and connecting the corresponding nodes to form a block of four cubes. To go to six dimensions, we could replicate the block of four cubes and interconnect the corresponding nodes, and so on. An n-dimensional cube formed this way is called a hypercube. Many parallel computers use this topology because the diameter grows linearly with the dimensionality. Put in other words, the diameter is the base 2 logarithm of the number of nodes, so, for example, a 10-dimensional hypercube has 1024 nodes but a diameter of only 10, giving excellent delay properties. Note that in contrast, 1024 nodes arranged as a 32 × 32 grid has a diameter of 62, more than six times worse than the hypercube. The price paid for the smaller diameter is that the fanout and thus the number of links (and the cost) is much larger for the hypercube. Nevertheless, the hypercube is a common choice for high-performance systems. Multicomputers come in all shapes and sizes, so it is hard to give a clean taxonomy of them. Nevertheless, two general ‘‘styles’’ stand out: the MPPs and the clusters. We will now study each of these in turn.

8.4.2 MPPs—Massively Parallel Processors The first category consists of the MPPs (Massively Parallel Processors), which are huge multimillion dollar supercomputers. These are used in science, in engineering, and in industry for very large calculations, for handling very large numbers of transactions per second, or for data warehousing (storing and managing immense databases). Initially, MPPs were primarily used as scientific supercomputers, but now most of them are used in commercial environments. In a sense, these machines are the successors to the mighty mainframes of the 1960s (but the connection is tenuous, sort of like a paleontologist claiming that a flock of sparrows is the successor to the Tyrannosaurus Rex). To a large extent, the MPPs have displaced SIMD machines, vector supercomputers, and array processors at the top of the digital food chain. Most of these machines use standard CPUs as their processors. Popular choices are the Intel Pentium, the Sun UltraSPARC, and the IBM PowerPC. What sets the MPPs apart is their use of a very high-performance proprietary interconnection network designed to move messages with low latency and at high

618

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

bandwidth. Both of these are important because the vast majority of all messages are small (well under 256 bytes), but most of the total traffic is caused by large messages (more than 8 KB). MPPs also come with extensive proprietary software and libraries. Another point that characterizes MPPs is their enormous I/O capacity. Problems big enough to warrant using MPPs invariably have massive amounts of data to be processed, often terabytes. These data must be distributed among many disks and need to be moved around the machine at great speed. Finally, another issue specific to MPPs is their attention to fault tolerance. With thousands of CPUs, several failures per week are inevitable. Having an 18hour run be aborted because one CPU crashed is unacceptable, especially when having one such failure is to be expected every week. Thus large MPPs always have special hardware and software for monitoring the system, detecting failures, and recovering from them smoothly. While it would be nice to study the general principles of MPP design now, in truth, there are not many principles. When you come right down to it, an MPP is a collection of more-or-less standard computing nodes connected by a very fast interconnect of the types we have already examined. So instead, we will now look at two examples of MPPs: BlueGene/L and Red Storm. BlueGene As a first example of a massively parallel processor, we will now examine the IBM BlueGene system. IBM conceived this project in 1999 as a massively parallel supercomputer for solving computationally-intensive problems in, among other fields, the life sciences. For example, biologists believe that the threedimensional structure of a protein determines its functionality, yet computing the 3D structure of one small protein from the laws of physics took years on the supercomputers of that period. The number of proteins found in human beings is over half a million and many of them are extremely large and their misfolding is known to be responsible for certain diseases (e.g., cystic fibrosis). Clearly, determining the 3D structure of all the human proteins would require increasing the world’s computing power by many orders of magnitude, and modeling protein folding is only one problem that BlueGene was designed to handle. Equally complex challenges in molecular dynamics, climate modeling, astronomy, and even financial modeling also require orders of magnitude improvement in supercomputing. IBM felt that there was enough of a market for massive supercomputing that it invested $100 million to design and build BlueGene. In Nov. 2001, Livermore National Laborary, run by the U.S. Dept. of Energy, signed up as a partner and first customer for the first version of the BlueGene family, called BlueGene/L. The goal of the BlueGene project was not just to produce the world’s fastest MPP, but to also to produce the most efficient one in terms of teraflops/dollar,

SEC. 8.4

619

MESSAGE-PASSING MULTICOMPUTERS

teraflops/watt, and teraflops/m 3 . For this reason, IBM rejected the philosophy behind previous MPPs, which was to use the fastest components money could buy. Instead, a decision was made to produce a custom system-on-a-chip component that was to run at a modest speed and low power in order to produce a very large machine with a high packing density. The first chip was delivered in June 2003. The first quarter of the BlueGene/L system, with 16,384 compute nodes, was fully operational by Nov. 2004, when it was certified as the fastest supercomputer on earth at 71 teraflops/sec. Running at a mere 0.4 megawatt, it also won the power efficiency race for its class at 177.5 megaflops/watt. The rest of the system, bringing the total to 65,536 compute nodes, was scheduled for delivery in the summer of 2005. The heart of the BlueGene/L system is the custom node chip illustrated in Fig. 8-36. It consists of two PowerPC 440 cores running at 700 MHz. The PowerPC 440 is a pipelined dual-issue superscalar processor popular in embedded systems. Each core has a pair of dual-issue floating-point units, which together can issue four floating-point instructions per clock cycle. The floating-point units have been augmented with a number of SIMD-type instructions sometimes useful in scientific computations on arrays. While no performance slouch, this chip is clearly not a top-of-the-line multiprocessor. North Compute CPU Floatingpoint unit

FPU

L1 caches

PowerPC 440 core

Custom chip

I D

Snooping

West FPU FPU

PowerPC 440 core

Interface to 3D-torus

L2 cache

I

SRAM

FPU

Up

4-MB L3 cache

East To card-level DDR SDRAM

L2 cache

D

Communication CPU South

Ethernet

Down

Figure 8-36. The BlueGene/L custom processor chip.

The two CPU cores on the chip are identical in structure but are intended to be programmed differently. One of them is for computing; the other is for handling communication among the 65,536 nodes.

620

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Three levels of cache are present on the chip. The first level consists of a split L1 cache with 32 KB for instructions and 32 KB for data. There is no coherency between the L1 caches on the two CPUs because the standard PowerPC 440 cores do not support that and it was decided to use unmodified cores. The second level is a unified cache consisting of a unified 2-KB cache. The L2 caches are really prefetch buffers rather than true caches. They snoop on each other and are cache consistent. The third level is a unified, consistent 4-MB shared cache that feeds both of the L2 caches. A memory reference that misses on the L1 cache but hits on the L2 cache takes about 11 clock cycles. A miss on L2 that hits on L3 takes about 28 cycles. Finally, a miss on L3 that has to go t o the main SDRAM takes about 75 cycles. A small SRAM is connected to the L2 caches. It is connected to the JTAG pins for booting, debugging, communicating with the main host, holding the system stack, and providing semaphore, barrier, and other synchronization operations. At the next level up, IBM designed a custom card that holds two of the chips shown in Fig. 8-36 along with 1 GB of RAM. Future versions of the card may hold as much as 4 GB. The chip and the card are shown in Fig. 8-37(a)–(b) respectively.

512-MB

512-MB SDRAM 512-MB SDRAM

Chip:

(a)

Card: 2 Chips 1 GB

(b)

Board Cabinet 16 Cards 32 Boards 32 Chips 512 Cards 16 GB 1024 Chips 512 GB (c)

(d)

System 64 Cabinets 2048 Boards 32,768 Cards 65,536 Chips 32 TB (e)

Figure 8-37. The BlueGene/L (a) chip. (b) card. (c) board. (d) cabinet. (e) system.

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

621

The cards are mounted on plug-in boards, with 16 cards per board for a total of 32 chips (and thus 32 compute CPUs) per board. Since each card contains 1 GB of SDRAM, the boards contain 16 GB apiece. One board is illustrated in Fig. 8-37(c). At the next level, 16 of these boards are plugged into the top of a midplane and another 16 are plugged into the bottom, filling a 60 cm × 90 cm cabinet with 1024 CPUs, an extremely high density. The two groups of 16 are separated by a switch that can remove one group from the system for maintenance while at the same time switching in a spare group. A cabinet is illustrated in Fig. 8-37(d). Finally, a full system, consisting of 64 cabinets with 65,536 compute CPUs and another 65,536 communication CPUs, is depicted in Fig. 8-37(e). With 131,072 dual-issue integer CPUs and 262,144 dual-issue floating-point CPUs, this system should be able to issue up to 786,432 instructions per clock cycle. However, one of the integer units feeds the floating-point units, reducing the instruction rate to 655,360 instructions per clock cycle, or 4.6 × 1014 instructions/sec, making it the biggest computer system ever built by a huge margin. The system is a multicomputer in the sense that no CPU has direct access to any memory except the 512 MB on its own card. No two CPUs shared the same memory. In addition, there is no demand paging because there are no local disks to page off. Instead, the system has 1024 I/O nodes, which are connected to disks and the other peripheral devices. All in all, while the system is extremely large, it is also quite straightforward with little new technology except in the area of high-density packaging. The decision to keep it simple was no accident since a major goal was high reliability and availability. Consequently, a great deal of careful engineering went into the power supplies, fans, cooling, and cabling with the goal of a mean-time-to-failure of 10 days. To connect all the chips, a scalable, high-performance interconnect is needed. The design used is a three-dimensional torus measuring 64 × 32 × 32. As a consequence, each CPU needs six connections, two to other CPUs logically above and below it, north and south of it, and east and west of it. These six connections are labeled east, west, north, south, up, and down, respectively in Fig. 8-36. Physically, each 1024-node cabinet is an 8 × 8 × 16 torus. Pairs of neighboring cabinets for an 8 × 8 × 32 torus. Four pairs of cabinets in the same row form an 8 × 32 × 32 torus. Finally, all 8 rows form a 64 × 32 × 32 torus. All links are thus point-to-point and operate at 1.4 Gbps. Since each of the 65,536 nodes has three links to ‘‘higher’’ numbered nodes, one in each dimension, the total bandwidth of the system is 275 terabits/sec. The information content of this book is about 300 million bits, including all the art in encapsulated PostScript format, so BlueGene/L could move 900,000 copies of this book per sec. Where they would go and who would want them is left as an exercise for the reader. Communication on the 3D torus is done in the form of virtual cut through routing. This technique is somehwat akin to store-and-forward packet switching,

622

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

except that entire packets are not stored before being forwarded. As soon as a byte has arrived at one node, it can be forwarded to the next one along the path, even before the entire packet has arrived. Both dynamic (adaptive) and deterministic (fixed) routing are possible. A small amount of special-purpose hardware on the chip is used to implement the virtual cut through. In addition to the main 3D torus used for data transport, four other communication networks are present. The second one is a combining network in the form of a tree. Many of the operations performed on highly-parallel systems such as BlueGene/L require the participation of all the nodes. For example, consider finding the minimum value of a set of 65,536 values, one held in each node. The combining network joins all the nodes in a tree and whenever two nodes send their respective values to a higher-level node, it selects out the smallest one and forwards it upward. In this way, far less traffic reaches the root than if all 65,636 nodes sent a message there. The third network is for global barriers and interrupts. Some algorithms work in phases with each node required to wait until all the others have completed the phase before starting the next phase. The barrier network allows the software to define these phases and provide a way to suspend all compute CPUs that reach the end of a phase until all of them have reached the end, at which time they are all released. Interrupts also use this network. The fourth and fifth networks both use gigabit Ethernet. One of them connects the I/O nodes to the file servers, which are external to BlueGene/L, and to the Internet beyond. The other one is used for debugging the system. Each of the compute and communication nodes runs a small, custom, lightweight kernel that supports a single user and a single process. This process has two threads, one running on each of the CPUs in the node. This simple structure was designed for high performance and high reliability. For additional reliability, application software can call a library procedure to make a checkpoint. Once all outstanding messages have been cleared from the network, a global checkpoint can be made and stored so that in the event of a system failure, the job can be restarted from the checkpoint, rather than from the beginning. The I/O nodes run a traditional Linux operating system and support multiple processes. For more information about BlueGene/L see (Adiga et al., 2002; Almasi et al., 2003a, 2003b; and Blumrich et al., 2005). Red Storm As our second example of an MPP, let us consider the Red Storm machine (also called Thor’s Hammer) at Sandia National Laboratory. Sandia is operated by Lockheed Martin and does classified and unclassified work for the U.S. Dept. of Energy. Some of the classified work concerns the design and simulation of nuclear weapons, which is highly compute intensive.

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

623

Sandia has been in this business for a long time and has had many leadingedge supercomputers over the years. For decades, it favored vector supercomputers, but eventually technology and economics made MPPs more cost effective. By 2002, the then-current MPP, called ASCI Red, was getting a bit creaky. Although it had 9460 nodes, collectively they had a mere 1.2 TB of RAM and 12.5 TB of disk space, and the system could barely crank out 3 teraflops/sec. So in the summer of 2002, Sandia selected Cray Research, a long-time supercomputer vendor, to build it a replacement for ASCI Red. The replacement was delivered in August 2004, a remarkably short design and implementation cycle for such a large machine. The reason it could be designed and delivered so quickly is that Red Storm uses almost entirely off-the-shelf parts, except for one custom chip used for routing. The CPU selected for Red Storm was the AMD Opteron. The Opteron has several key characteristics that made it the first choice. The first one is that it has three operating modes. In legacy mode, it runs standard Pentium binary programs unmodified. In compatibility mode, the operating system runs in 64-bit mode and can address 264 bytes of memory, but application programs run in 32-bit mode. Finally, in 64-bit mode, the entire machine is 64 bits and all programs can address the full 64-bit address space. In 64-bit mode, it is possible to mix and match software: both 32-bit and 64-bit programs can run at the same time, allowing an easy upgrade path. The second key characteristic the Opteron has is its attention to the memory bandwidth problem. In recent years, CPUs have been getting faster and faster and memory has not been keeping pace, resulting in a massive penalty when there is a level 2 cache miss. AMD integrated the memory controller into the Opteron so it can run at the speed of the processor clock instead of the speed of the memory bus, which improves memory performance. The controller can handle eight DIMMS of 4 GB each, for a maximum total memory of 32 GB per Opteron. In the Red Storm system, each Opteron has only 2–4 GB. However, as memory gets cheaper, no doubt more will be added in the future though. An upgrade to dualcore Opterons is also an option, doubling the raw compute power. Each Opteron has its own dedicated custom network processor called the Seastar, manufactured by IBM. The Seastar is a critical component since nearly all the data traffic between the processors goes over the Seastar network. Without the very high-speed interconnect provided by these custom chips, the system would quickly bog down in data. Although the Opterons are commercially available off the shelf, the Red Storm packaging is custom-built. Each Red Storm board contains four Opterons, 4 GB of RAM, four Seastars, a RAS (Reliability, Availability, and Service) processor, and a 100-Mbps Ethernet chip, as shown in Fig. 8-38. A set of eight boards is plugged into a backplane and inserted into a card cage. Each cabinet holds three card cages for a total of 96 Opterons, plus the necessary power supplies and fans. The full system consists of 108 cabinets for

624

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

2–4 GB SDRAM Set of 8 boards

Card cage

M

M

M

M

CPU

CPU

CPU

CPU

RAS

Card cage

Opteron 100-Mbps Ethernet

Seastar

Card cage

Cabinet

Figure 8-38. Packaging of the Red Storm components.

compute nodes, giving a total of 10,368 Opterons with 10 TB of SDRAM. Each CPU has access to only its own SDRAM. There is no shared memory. The theoretical computing power of the system is 41 teraflops/sec. The interconnection between the Opteron CPUs is done by the custom Seastar routers, one router per Opteron CPU. The are connected in a 3D torus of size 27 × 16 × 24 with one Seastar at each mesh point. Each Seastar has seven bidirectional 24-Gbps links, going north, east, south, west, up, down, and to the Opteron. The transit time beween adjacent mesh points is 2 microsec. Across the entire set of compute nodes it is only 5 microsec. A second network using 100Mbps Ethernet is used for service and maintenance. In addition to the 108 compute cabinets, the system also contains 16 cabinets for I/O and service processors. Each of these cabinets holds 32 Opterons. These 512 CPUs are split: 256 for I/O and 256 for service. The rest of the space is for disks, which are organized as RAID 3 and RAID 5, each with a parity drive and a hot spare. The total disk space is 240 TB. The combined sustained disk bandwidth is 50 GB/sec. The system is partitioned into classified and unclassified sections, with switches between the parts so they can be mechanically coupled or decoupled. A

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

625

total of 2688 compute CPUs are always in the classified section and another compute 2688 are always in the unclassified section. The remaining 4992 compute CPUs are switchable into either section, as depicted in Fig. 8-39. The 2688 classified Opterons each have 4 GB of RAM; all the rest have 2 GB each. The I/O and service processors are split between the two parts. I/O and service node

28 Classified 120 TB cabinets storage (2688 Opterons

Switch

Compute node

52 Switchable cabinets 28 Unclassified 120 TB (4992 Opterons ) cabinets (2688 Opterons) storage )

Figure 8-39. The Red Storm system as viewed from above.

Everything is housed in a new 2000 m2 custom building. The building and site have been designed so that the system can be upgraded to as many as 30,000 CPUs in the future if required. The compute nodes draw 1.6 megawatts of power; the disks draw another megawatt. Adding in the fans and air conditioning, the whole thing uses 3.5 MW. The computer hardware and software cost $90 million. The building and cooling cost another $9 million, so the total cost came in at just under $100 million, although some of that was nonrecurring engineering cost. If you want to order an exact clone, $60 million would be a good number to keep in mind. Cray intends to sell smaller versions of the system to other government and commercial customers under the name X3T. The compute nodes run a lightweight kernel called catamount. The I/O and service nodes run plain vanilla Linux with a small addition to support MPI (discussed later in this chapter). The RAS nodes run a stripped down Linux. Extensive software from ASCI Red is available for use on Red Storm, including CPU allocators, schedulers, MPI libraries, math libraries, as well as the application programs. With such a large system, achieving high reliability is essential. Each board has a RAS processor for doing system maintenance and there are special hardware facilities as well. The goal is an MTBF (Mean Time Between Failures) of 50 hours. ASCI Red had a hardware MTBF of 900 hours, but was plagued by an operating system crash every 40 hours. Although the new hardware is much more reliable than the old hardware, the weak point remains the software. For more information about Red Storm, see (Brightwell et al., 2005).

626

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

A Comparison of BlueGene/L and Red Storm Red Storm and BlueGene/L are comparable in some ways but different in other ways, so it is interesting to put some of the key parameters next to each other, as presented in Fig. 8-40. ################################################################# ! ! ! Item BlueGene/L Red Storm !################################################################# ! ! ! CPU !################################################################# ! 32-Bit PowerPC ! 64-Bit Opteron ! ! Clock ! 700 MHz ! 2 GHz ! !################################################################# ! ! ! !################################################################# ! 65,536 ! 10,368 ! Compute CPUs ! ! ! ! CPUs/board !################################################################# ! 32 ! 4 ! !################################################################# ! 1024 ! 96 ! CPUs/cabinet ! ! ! ! Compute cabinets ! 64 !################################################################# ! 108 ! ! Teraflops/sec ! 71 ! 41 ! ################################################################# ! ! ! ! Memory/CPU !################################################################# ! 512 MB ! 2–4 GB ! ! Total memory ! 32 TB ! 10 TB ! !################################################################# ! ! ! !################################################################# ! PowerPC ! Seastar ! Router ! ! ! ! Number of routers ! 65,536 !################################################################# ! 10,368 ! ! Interconnect ! 3D torus 64 × 32 × 32 ! 3D torus 27 × 16 × 24 ! ################################################################# ! ! ! ! ! Other networks ! Gigabit Ethernet ! Fast Ethernet ! ################################################################# ! Partitionable ! No ! Yes ! ################################################################# ! ! ! ! Compute OS !################################################################# ! Custom ! Custom ! ! ! ! ! I/O OS !################################################################# ! Linux ! Linux ! !################################################################# ! IBM ! Cray Research ! Vendor ! ! ! ! ! Expensive ################################################################# ! Yes ! Yes ! !

Figure 8-40. A comparison of BlueGene/L and Red Storm.

The two machines were built in the same time frame, so their differences are not due to technology but to designers’ different visions and to some extent the differences between the builders, IBM and Cray. BlueGene/L was designed from the beginning as a commercial machine of which IBM hopes to sell large numbers to biotech, pharmaceutical, and other companies. Red Storm was built on special contract with Sandia, although Cray plans to make a smaller version for sale, too. IBM’s vision is clear: combine existing cores to produce a custom chip that can be mass produced cheaply, run it at a low speed, and hook together a very large number of them using a modest speed communication network. Sandia’s vision is equally clear, but different: use a powerful off-the-shelf 64-bit CPU, design a very fast custom router chip, and throw in a lot of memory to produce a far more powerful node than BlueGene/L so fewer will be needed and communication between them will be faster.

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

627

The consequences of these decisions had consequences for the packaging. Because IBM built a custom chip combining the processor and router, it achieved a higher packing density: 1024 CPUs/cabinet. Because Sandia went for an unmodified off-the-shelf CPU chip and 2–4 GB of RAM per node, it could only get 96 compute CPUs in a cabinet. The result is that Red Storm takes up more floor space and consumes more power than BlueGene/L. In the exotic world of national laboratory computing, the bottom line is performance. In this respect, BlueGene/L wins, 71 TF/sec to 41 TF/sec, but Red Storm was designed to be expandable, so by throwing another 10,368 Opterons at the problem (for example, by switching to dual-core chips), Sandia could probably up its performance to 82 TF. IBM could respond by cranking the clock up a bit (700 MHz is not really pushing the state-of-the-art very hard). In short, MPP supercomputers have not even come close to any physical limits yet and will continue growing for years to come.

8.4.3 Cluster Computing The other style of multicomputer is the cluster computer (Anderson et al., 1995; Martin et al., 1997). It typically consists of hundreds or thousands of PCs or workstations connected by a commercially-available network board. The difference between an MPP and a cluster is analogous to the difference between a mainframe and a PC. Both have a CPU, both have RAM, both have disks, both have an operating system, and so on. The mainframe just has faster ones (except maybe the operating system). Yet qualitatively they feel different and are used and managed differently. This same difference holds for MPPs versus clusters. Historically, the key element that made MPPs special was their high-speed interconnect, but the recent arrival of commercial, off-the-shelf, high-speed interconnects has begun to close the gap. All in all, clusters are likely to drive MPPs into ever tinier niches, just as PCs have turned mainframes into esoteric specialty items. The main niche for MPPs is high-budget supercomputers, where peak performance is everything and if you have to ask the price you cannot afford one. While many kinds of clusters exist, two species dominate: centralized and decentralized. A centralized cluster is a cluster of workstations or PCs mounted in a big rack in a single room. Sometimes they are packaged in a much more compact way than usual to reduce physical size and cable length. Typically, the machines are homogeneous and have no peripherals other than network cards and possibly disks. Gordon Bell, the designer of the PDP-11 and VAX, has called such machines headless workstations (because they have no owners). We were tempted to call them headless COWs, but feared such a term would gore too many holy cows, so we refrained. Decentralized clusters consist of the workstations or PCs spread around a building or campus. Most of them are idle many hours a day, especially at night. Usually, these are connected by a LAN. Typically, they are heterogeneous and

628

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

have a full complement of peripherals, although having a cluster with 1024 mice is really not much better than a cluster with 0 mice. Most importantly, many of them have owners who have emotional attachments to their machines and tend to frown upon some astronomer trying to simulate the big bang on theirs. Using idle workstations to form a cluster invariably means having some way to migrate jobs off machines when their owners want to reclaim them. Job migration is possible but adds software complexity. Clusters are often smallish affairs, ranging from a dozen to perhaps 500 PCs. However, it is also possible to build very large ones from off-the-shelf PCs. Google has done this in an interesting way that we will now look at. Google Google is a popular search engine for finding information on the Internet. While its popularity is due, in part, to its simple interface and fast response time, its design is anything but simple. From Google’s point of view, the problem is that it has to find, index, and store the entire World Wide Web (over 8 billion pages and 1 billion images), be able to search the whole thing in under 0.5 sec, and handle thousands of queries/sec coming from all over the world 24 hours a day. In addition, it must never go down, not even in the face of earthquakes, electrical power failures, telecom outages, hardware failures and software bugs. And of course, it has to do all of this as cheaply as possible. Building a Google clone is definitely not an exercise for the reader. How does Google do it? To start with, Google operates multiple data centers around the world. Not only does this approach provide backups in case one of them is swallowed by an earthquake, but when www.google.com is looked up, the sender’s IP address is inspected and the address of the nearest data center is supplied. The browser then sends the query there. Each data center has at least one OC-48 (2.488 Gbps) fiber optics connection to the Internet, on which it receives queries and sends answers, as well as an OC12 (622 Mbps) backup connection from a different telecom provider, in case the primary ones go down. Uninterruptable power supplies and emergency diesel generators are available at all data centers to keep the show going during power failures. Consequently, during a major natural disaster, performance will suffer, but Google will keep running. To get a better understanding of why Google chose the architecture it did, it is useful to briefly describe how a query is processed once it hits its designated data center. After arriving at the data center (step 1 in Fig. 8-41), the load balancer routes the query to one of the many query handlers (2), and to the spelling checker (3) and ad server (4) in parallel. Then the search words are looked up on the index servers (5) in parallel. These servers contain an entry for each word on the Web. Each entry lists all the documents (Web pages, PDF files, PowerPoint presentations, etc.) containing the word, sorted in page rank order. Page rank is

SEC. 8.4

629

MESSAGE-PASSING MULTICOMPUTERS

determined by a complicated (and secret) formula, but the number of links to a page and their own ranks play a large role. To get higher performance, the index is divided into pieces called shards that can be searched in parallel. Conceptually, at least, shard 1 contains all the words in the index, with each one followed by the IDs of the n highest-ranked documents containing that word. Shard 2 contains all the words and the IDs of the n next highest documents, and so on. As the Web grows, each of these shards may later be split with the first k words in one set of shards, the next k words in a second set of shards and so forth, in order to achieve even more search parallelism. 1 11 Load balancer 2

3 Spell checker

aardvark abacus abalone abandon

9

4

Query handler 5

7

6

154, 3016, ... 973, 67231, ... 73403,89021, ... 14783, 63495, ... . . .

10 8

5

8

Index servers

Ad server

Document servers

6 aardvark abacus abalone abandon

1242, 5643 ... 8393, 65837, ... 59343, 93082, ... 40323, 94834, ... . . .

7

Figure 8-41. Processing of a Google query.

The index servers return a set of document identifiers (6) that are then combined according to the Boolean properties of the query. For example, if the search was for +digital +capybara +dance, then only document identifiers appearing in all three sets are used in the next step. In this step (7), the documents themselves are referenced to extract their titles, URLs, and snippets of text surrounding the search terms. The document servers contain many copies of the entire Web at each data center, hundreds of terabytes at present. The documents are also divided into shards to enhance parallel search. While processing a query does not require reading the whole Web (or even reading the tens of terabytes on the index servers), having to process 100 MB per query is normal.

630

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

When the results are returned to the query handler (8), the pages found are collated into page rank order. If potential spelling errors are detected (9), they are announced and relevant ads are added (10). Displaying ads for advertisers interested in buying specific search terms (e.g., ‘‘hotel’’ or ‘‘camcorder’’) is how Google makes its money. Finally, the results are formatted in HTML (HyperText Markup Language) and sent to the user as a Web page. With this background, we can now examine the Google architecture. Most companies, when faced with a huge data base, massive transaction rate, and the need for high reliability, would buy the biggest, fastest, and most reliable equipment on the market. Google did just the opposite. It bought cheap, modestperformance PCs. Lots of them. And with them, it built the world’s largest offthe-shelf cluster. The driving principle behind this decision was simple: Optimize price/performance. The logic behind this decision lies in economics: commodity PCs are very cheap. High-end servers are not and large multiprocessors are even less so. Thus while a high-end server might have two or three times the performance of a midrange desktop PC, it will typically be 5–10 times the price, which is not cost effective. Of course, cheap PCs fail more often than top-of-the-line servers, but the latter fail, too, so the Google software had to be designed to work with failing hardware no matter what kind of equipment it was using. Once the fault-tolerance software was written, it did not really matter whether the failure rate was 0.5% per year or 2% per year, failures had to be dealt with. Google’s experience is that about 2% of the PCs fail each year. More than half the failures are due to faulty disks, followed by power supplies and then RAM chips. After burn in, CPUs never fail. Actually, the biggest source of crashes is not hardware at all; it is software. The first response to a crash is just to reboot, which often solves the problem (the electronic equivalent of the doctor saying: ‘‘Take two aspirins and go to bed.’’). A typical modern Google PC consists of a 2-GHz Pentium, 512 MB of RAM, and a disk of around 80 GB, the kind of thing a grandmother might buy for checking her e-mail occasionally. The only specialty item is an Ethernet chip. Not exactly state of the art, but very cheap. The PCs are housed in 1u-high cases (about 5 cm thick) and stacked 40 high in 19-inch racks, one stack in front and one stack in back for a total of 80 PCs per rack. The PCs in a rack are connected by switched Ethernet, with the switch inside the rack. The racks in a data center are also connected by switched Ethernet, with two redundant switches per data center used to survive switch failures. The layout of a typical Google data center is illustrated in Fig. 8-42. The incoming high-bandwidth OC-48 fiber is routed to each of two 128-port Ethernet switches. Similarly, the backup OC-12 fiber is also routed to each of the two switches. The incoming fibers use special input cards and do not occupy any of the 128 Ethernet ports. Each rack has four Ethernet links coming out of it, two to

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

631

the left switch and two to the right switch. In this configuration, the system can survive the failure of either switch. Since each rack has four connections to the switch (two from the front 40 PCs and two from the back 40 PCs), it takes four link failures or two link failures and a switch failure to take a rack offline. With a pair of 128-port switches and four links from each rack, up to 64 racks can be supported. With 80 PCs per rack, a data center can have up to 5120 PCs. But, of course, racks do not have to hold exactly 80 PCs and switches can be larger or smaller than 128 ports; these are just typical values for a Google cluster. OC-12 Fiber

128-port Gigabit Ethernet switch

OC-48 Fiber

128-port Gigabit Ethernet switch

Two gigabit Ethernet links

80-PC rack

Figure 8-42. A typical Google cluster.

Power density is also a key issue. A typical PC burns about 120 watts or about 10 kW per rack. A rack needs about 3 m2 so maintenance personnel can install and remove PCs and for the air conditioning to function. These parameters give a power density of over 3000 watts/m 2 . Most data centers are designed for 600–1200 watts/m 2 , so special measures are required to cool the racks. Google has learned three key things about running massive Web servers that bear repeating.

632

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

1. Components will fail so plan for it. 2. Replicate everything for throughput and availability. 3. Optimize price/performance. The first item says that you need to have fault-tolerant software. Even with the best of equipment, when you have a massive number of components, some will fail and the software has to be able to handle it. Whether you have 1 failure a week or two failures a week, the software has to be able to handle failures with systems of this size. The second item points out that both hardware and software have to be highly redundant. Doing so not only improves the fault-tolerance properties, but also the throughput. In the case of Google, the PCs, disks, cables, and switches are all replicated many times over. Furthermore, the index and the documents are broken into shards and the shards are heavily replicated in each data center and the data centers are themselves replicated. The third item is a consequence of the first two. If the system has been properly designed to deal with failures, buying expensive components such as RAIDs with SCSI disks is a mistake. Even they will fail, but spending 10 times as much to cut the failure rate in half is a bad idea. Better to buy 10 times as much hardware and deal with the failures when they occur. At the very least, having more hardware will give better performance when everything is working. For more information about Google, see (Barroso et al., 2003; and Ghemawat et al., 2003).

8.4.4 Communication Software for Multicomputers Programming a multicomputer requires special software, usually libraries, for handling interprocess communication and synchronization. In this section we will say a few words about this software. For the most part, the same software packages run on MPPs and clusters, so applications can be easily ported between platforms. Message-passing systems have two or more processes running independently of one another. For example, one process may be producing some data and one or more others may be consuming it. There is no guarantee that when the sender has more data the receiver(s) will be ready for it, as each one runs its own program. Most message-passing systems provide two primitives (usually library calls), send and receive, but several different kinds of semantics are possible. The three main variants are 1. Synchronous message passing. 2. Buffered message passing. 3. Nonblocking message passing.

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

633

In synchronous message passing, if the sender executes a send and the receiver has not yet executed a receive, the sender is blocked (suspended) until the receiver executes a receive, at which time the message is copied. When the sender gets control back after the call, it knows that the message has been sent and correctly received. This method has the simplest semantics and does not require any buffering. It has the severe disadvantage that the sender remains blocked until the receiver has gotten and acknowledged receipt of the message. In buffered message passing, when a message is sent before the receiver is ready, the message is buffered somewhere, for example, in a mailbox, until the receiver takes it out. Thus in buffered message passing, a sender can continue after a send, even if the receiver is busy with something else. Since the message has actually been sent, the sender is free to reuse the message buffer immediately. This scheme reduces the time the sender has to wait. Basically, as soon as the system has sent the message the sender can continue. However, the sender now has no guarantee that the message was correctly received. Even if communication is reliable, the receiver may have crashed before getting the message. In nonblocking message passing, the sender is allowed to continue immediately after making the call. All the library does is tell the operating system to do the call later, when it has time. As a consequence, the sender is hardly blocked at all. The disadvantage of this method is that when the sender continues after the send, it may not reuse the message buffer as the message may not yet have been sent. Somehow it has to find out when it can reuse the buffer. One idea is to have it poll the system to ask. The other is to get an interrupt when the buffer is available. Neither of these makes the software any simpler. Below we will briefly discuss a popular message-passing systems available on many multicomputers: MPI. MPI—Message-Passing Interface For quite a few years, the most popular communication package for multicomputers was PVM Parallel Virtual Machine (Geist et al., 1994; and Sunderram, 1990). However, in recent years it has been largely replaced by MPI (Message-Passing Interface). MPI is much richer and more complex than PVM, with many more library calls, many more options, and many more parameters per call. The original version of MPI, now called MPI-1, was augmented by MPI-2 in 1997. Below we will give a very cursory introduction to MPI-1 (which contains all the basics), then say a little about what was added in MPI-2. For more information about MPI, see (Gropp et al., 1994; and Snir et al., 1996). MPI-1 does not deal with process creation or management, as PVM does. It is up to the user to create processes using local system calls. Once they have been created, they are arranged into static, unchanging process groups. It is with these groups that MPI works.

634

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

MPI is based on four major concepts: communicators, message data types, communication operations, and virtual topologies. A communicator is a process group plus a context. A context is a label that identifies something, such as a phase of execution. When messages are sent and received, the context can be used to keep unrelated messages from interfering with one another. Messages are typed and many data types are supported, including characters, short, regular, and long integers, single- and double-precision floating-point numbers, and so on. It is also possible to construct other types derived from these. MPI supports an extensive set of communication operations. The most basic one is used to send messages as follows: MPI"Send(buffer, count, data"type, destination, tag, communicator)

This call sends a buffer with count number of items of the specified data type to the destination. The tag field labels the message so the receiver can say it only wants to receive a message with that tag. The last field tells which process group the destination is in (the destination field is just an index into the list of processes for the specified group). The corresponding call for receiving a message is MPI"Recv(&buffer, count, data"type, source, tag, communicator, &status)

which announces that the receiver is looking for a message of a certain type from a certain source with a certain tag. MPI supports four basic communication modes. Mode 1 is synchronous, in which the sender may not begin sending until the receiver has called MPI"Recv. Mode 2 is buffered, in which this restriction does not hold. Mode 3 is standard, which is implementation dependent and can be either synchronous or buffered. Mode 4 is ready, in which the sender claims the receiver is available (as in synchronous), but no check is made. Each of these primitives comes in a blocking and a nonblocking version, leading to eight primitives in all. Receiving only has two variants: blocking and nonblocking. MPI supports collective communication, including broadcast, scatter/gather, total exchange, aggregation, and barrier. For all forms of collective communication, all the processes in a group must make the call and with compatible arguments. Failure to do this is an error. A typical form of collective communication is for processes organized in a tree, in which values propagate up from the leaves to the root, undergoing some processing at each step, for example, adding up the values or taking the maximum. The fourth basic concept in MPI is the virtual topology, in which the processes can be arranged in a tree, ring, grid, torus, or other topology. Such an arrangement provides a way to name communication paths and facilitates communication. MPI-2 adds dynamic processes, remote memory access, nonblocking collective communication, scalable I/O support, real-time processing, and many other new features that are beyond the scope of this book. In the scientific community,

SEC. 8.4

635

MESSAGE-PASSING MULTICOMPUTERS

a battle raged for years between the MPI and PVM camps. The PVM side said that PVM was easier to learn and simpler to use. The MPI side said the MPI does more and also points out that MPI is a formal standard with a standardization committee and an official defining document. The PVM side agreed but claimed the lack of a full-blown standardization bureaucracy is not necessarily a drawback. When all was said and done, it appears that MPI won.

8.4.5 Scheduling MPI programmers can easily create jobs requesting multiple CPUs and running for substantial periods of time. When multiple independent requests are available from different users, each needing a different number of CPUs for different time periods, the cluster needs a scheduler to determine which job gets to run when. In the simplest model, the job scheduler requires each job to specify how many CPUs it needs. Jobs are then run in strict FIFO order, as shown in Fig. 843(a). In this model, after a job is started, a check is made to see if enough CPUs are available to start the next job in the input queue. If so, it is started, and so on. Otherwise, the system waits until more CPUs become available. As an aside, although we have suggested that this cluster has eight CPUs, it might well have 128 CPUs that are allocated in units of 16 (giving eight CPU groups), or some other combination. CPU group

CPU group

CPU group

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7

0 1 2 3 4 5 6 7

1

1 2

6

Time

6 8

9

7

3 6

9

5 5

4

7

3

3

4

1

4

5

8

2 9

2

8

7

(a)

(b)

(c)

Figure 8-43. Scheduling a cluster. (a) FIFO. (b) Without head-of-line blocking. (c) Tiling. The shaded areas indicate idle CPUs.

A better scheduling algorithm avoids head-of-line blocking by skipping over jobs that do not fit and picking the first one that does fit. Whenever a job finishes,

636

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

the queue of remaining jobs is checked in FIFO order. This algorithm gives the result of Fig. 8-43(b). A still more sophisticated scheduling algorithm requires each submitted job to specify its shape, that is, how many CPUs for how many minutes. With that information, the job scheduler can attempt to tile the CPU-time rectangle. Tiling is especially effective when jobs are submitted during the daytime for execution at night, so the job scheduler has all the information about all the jobs in advance and can run them in optimal order, as illustrated in Fig. 8-43(c).

8.4.6 Application-Level Shared Memory That multicomputers scale to larger sizes than multiprocessors should be clear from our examples. This reality has led to the development of message-passing systems like MPI. Many programmers do not like this model and would like to have the illusion of shared memory, even if it is not really there. Achieving this goal would be the best of both worlds: large, inexpensive hardware (at least, per node) plus ease of programming. This is the holy grail of parallel computing. Many researchers have concluded that while shared memory at the architectural level may not scale well, there may be other ways to achieve the same goal. From Fig. 8-19, we see that there are other levels at which a shared memory can be introduced. In the following sections, we will look at some ways that shared memory can be introduced into the programming model on a multicomputer, without it being present at the hardware level. Distributed Shared Memory One class of application-level shared memory system is the page-based system. It goes under the name of DSM (Distributed Shared Memory). The idea is simple: a collection of CPUs on a multicomputer share a common paged virtual address space. In the simplest version, each page is held in the RAM of exactly one CPU. In Fig. 8-44(a), we see a shared virtual address space consisting of 16 pages, spread over four CPUs. When a CPU references a page in its own local RAM, the read or write just happens without any further delay. However, when a CPU references a page in a remote memory, it gets a page fault. Only instead of having the missing page being brought in from disk, the runtime system or operating system sends a message to the node holding the page to unmap it and send it over. After it has arrived, it is mapped in and the faulting instruction restarted, just as with a normal page fault. In Fig. 8-44(b), we see the situation after CPU 0 has faulted on page 10: it is moved from CPU 1 to CPU 0. This basic idea was first implemented in IVY (Li and Hudak, 1986, 1989). It provides a fully shared, sequentially consistent memory on a multicomputer. However, many optimizations are possible to improve the performance. The first

SEC. 8.4

637

MESSAGE-PASSING MULTICOMPUTERS Globally shared virtual memory consisting of 16 pages

0

0

1

2

2

5

9

3

4

5

6

1

3

8

10

7

8

6

9

4

7

12

14

CPU 1

CPU 0

10 11 12 13 14 15

2

9

10

5

1

3

6

8

CPU 0

2

CPU 3 Network

4

7

12

14

CPU 1

9

10 CPU 0

5

1

3

11

8

10

13

CPU 2

6

CPU 1

15

CPU 3 Network

4

7

12

14 CPU 2

(c)

15 Memory

(b)

0

13

CPU 2

(a)

0

11

11

13

15

CPU 3 Network

Figure 8-44. A virtual address space consisting of 16 pages spread over four nodes of a multicomputer. (a) The initial situation. (b) After CPU 0 references page 10. (c) After CPU 1 references page 10, here assumed to be a read-only page.

optimization, present in IVY, is to allow pages that are marked as read-only to be present at multiple nodes at the same time. Thus when a page fault occurs, a copy of the page is sent to the faulting machine, but the original stays where it is since there is no danger of conflicts. The situation of two CPUs sharing a read-only page (page 10) is illustrated in Fig. 8-44(c). Even with this optimization, performance is frequently unacceptable, especially when one process is actively writing a few words at the top of some page

638

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

and another process on a different CPU is actively writing a few words at the bottom of the page. Since only one copy of the page exists, the page must be pingponged back and forth constantly, a situation known as false sharing. The problem of false sharing can be attacked in several ways. In the Treadmarks system, for example, sequentially consistent memory is abandoned in favor of release consistency (Amza, 1996). Potentially writable pages may be present at multiple nodes at the same time, but before doing a write, a process must first do an acquire operation to signal its intention. At that point, all copies but the most recent one are invalidated. No other copies may be made until the corresponding release is done, at which time the page can be shared again. A second optimization done in Treadmarks is to initially map each writable page in read-only mode. When the page is first written to, a protection fault occurs and the system makes a copy of the page, called the twin. Then the original page is mapped in as read-write and subsequent writes can go at full speed. When a remote page fault happens later and the page has to be shipped over there, a word-by-word comparison is done between the current page and the twin. Only those words that have been changed are sent, reducing the size of the messages. When a page fault occurs, the missing page has to be located. Various solutions are possible, including those used in NUMA and COMA machines, such as (home-based) directories. In fact, many of the solutions used in DSM are also applicable to NUMA and COMA because DSM is really just a software implementation of NUMA or COMA with each page being treated like a cache line. DSM is a hot area of research. Interesting systems include CASHMERE (Kontothanassis, et al., 1997; and Stets et al., 1997), CRL (Johnson et al., 1995), Shasta (Scales et al., 1996), and Treadmarks (Amza, 1996; and Lu et al., 1997). Linda Page-based DSM systems like IVY and Treadmarks use the MMU hardware to trap accesses to missing pages. While making and sending differences instead of whole pages helps, the fact remains that pages are an unnatural unit for sharing, so other approaches have been tried. One such approach is Linda, which provides processes on multiple machines with a highly structured distributed shared memory (Carriero and Gelernter, 1989). This memory is accessed through a small set of primitive operations that can be added to existing languages, such as C and FORTRAN, to form parallel languages, in this case, C-Linda and FORTRAN-Linda. The unifying concept behind Linda is that of an abstract tuple space, which is global to the entire system and accessible to all processes in it. Tuple space is like a global shared memory, only with a certain built-in structure. The tuple space contains some number of tuples, each of which consists of one or more fields. For C-Linda, field types include integers, long integers, and floating-point num-

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

639

bers, as well as composite types such as arrays (including strings) and structures (but not other tuples). Figure 8-45 shows three tuples as examples. (′′abc′′, 2, 5) (′′matrix-1′′, 1, 6, 3.14) (′′family′′, ′′is sister′′, Carolyn, Elinor) Figure 8-45. Three Linda tuples.

Four operations are provided on tuples. The first one, out, puts a tuple into the tuple space. For example, out(′′abc′′, 2, 5);

puts the tuple (′′abc′′, 2, 5) into the tuple space. The fields of out are normally constants, variables, or expressions, as in out(′′matrix−1′′, i, j, 3.14);

which outputs a tuple with four fields, the second and third of which are determined by the current values of the variables i and j. Tuples are retrieved from the tuple space by the in primitive. They are addressed by content rather than by name or address. The fields of in can be expressions or formal parameters. Consider, for example, in(′′abc′′, 2, ? i);

This operation ‘‘searches’’ the tuple space for a tuple consisting of the string ′′abc′′, the integer, 2, and a third field containing any integer (assuming that i is an integer). If found, the tuple is removed from the tuple space and the variable i is assigned the value of the third field. The matching and removal are atomic, so if two processes execute the same in operation simultaneously, only one of them will succeed, unless two or more matching tuples are present. The tuple space may even contain multiple copies of the same tuple. The matching algorithm used by in is straightforward. The fields of the in primitive, called the template, are (conceptually) compared to the corresponding fields of every tuple in the tuple space. A match occurs if the following three conditions are all met: 1. The template and the tuple have the same number of fields. 2. The types of the corresponding fields are equal. 3. Each constant or variable in the template matches its tuple field. Formal parameters, indicated by a question mark followed by a variable name or type, do not participate in the matching (except for type checking), although those containing a variable name are assigned after a successful match.

640

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

If no matching tuple is present, the calling process is suspended until another process inserts the needed tuple, at which time the caller is automatically revived and given the new tuple. The fact that processes block and unblock automatically means that if one process is about to output a tuple and another is about to input it, it does not matter which goes first. In addition to out and in, Linda also has a primitive read, which is the same as in except that it does not remove the tuple from the tuple space. There is also a primitive eval, which causes its parameters to be evaluated in parallel and the resulting tuple to be deposited in the tuple space. This mechanism can be used to perform an arbitrary computation. This is how parallel processes are created in Linda. A common programming paradigm in Linda is the replicated worker model. This model is based on the idea of a task bag full of jobs to be done. The main process starts out by executing a loop containing out(′′task-bag′′, job);

in which a different job description is output to the tuple space on each iteration. Each worker starts out by getting a job description tuple using in(′′task-bag′′, ?job);

which it then carries out. When it is done, it gets another. New work may also be put into the task bag during execution. In this simple way, work is dynamically divided among the workers, and each worker is kept busy all the time, all with relatively little overhead. Various implementations of Linda on multicomputer systems exist. In all of them, a key issue is how to distribute the tuples among the machines and how to locate them when needed. Various possibilities include broadcasting and directories. Replication is also an important issue. These points are discussed in Bjornson (1993). Orca A somewhat different approach to application-level shared memory on a multicomputer is to use full-blown objects instead of just tuples as the unit of sharing. Objects consist of internal (hidden) state plus methods for operating on that state. By not allowing the programmer to access the state directly, many possibilities are opened to allow sharing over machines that do not have physical shared memory. One object-based system that gives the illusion of shared memory on multicomputer systems is Orca (Bal, 1991; Bal et al., 1992; and Bal and Tanenbaum, 1988). Orca is a traditional programming language (based on Modula 2) to which two new features have been added: objects and the ability to create new processes. An Orca object is an abstract data type, analogous to an object in Java or a package in Ada. It encapsulates internal data structures and user-written methods,

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

641

called operations. Objects are passive, that is, they do not contain threads to which messages can be sent. Instead, processes access an object’s internal data by invoking its methods. Each Orca method consists of a list of (guard, block-of-statements) pairs. A guard is a Boolean expression that does not contain any side effects, or the empty guard, which is the same as the value true. When an operation is invoked, all of its guards are evaluated in an unspecified order. If all of them are false, the invoking process is delayed until one becomes true. When a guard is found that evaluates to true, the block of statements following it is executed. Figure 8-46 depicts a stack object with two operations, push and pop. Object implementation stack; top:integer; stack: array [integer 0..N−1] of integer;

# storage for the stack

operation push(item: integer); begin guard top < N − 1 do stack[top] := item; top := top + 1; od; end;

# function returning nothing

operation pop( ): integer; begin guard top > 0 do top := top − 1; return stack[top]; od; end;

# function returning an integer

begin top := 0; end;

# push item onto the stack # increment the stack pointer

# suspend if the stack is empty # decrement the stack pointer # return the top item

# initialization

Figure 8-46. A simplified ORCA stack object, with internal data and two operations.

Once a stack has been defined, variables of this type can be declared, as in s, t: stack;

which creates two stack objects and initializes the top variable in each to 0. The integer variable k can be pushed onto the stack s by the statement s$push(k);

and so forth. The pop operation has a guard, so an attempt to pop a variable from an empty stack will suspend the caller until another process has pushed something on the stack.

642

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

Orca has a fork statement to create a new process on a user-specified processor. The new process runs the procedure named in the fork statement. Parameters, including objects, may be passed to the new process, which is how objects become distributed among machines. For example, the statement for i in 1 .. n do fork foobar(s) on i; od; generates one new process on each of machines 1 through n, running the program foobar in each of them. As these n new processes (and the parent) execute in parallel, they can all push and pop items onto the shared stack s as though they were all running on a shared-memory multiprocessor. It is the job of the runtime system to sustain the illusion of shared memory where it really does not exist. Operations on shared objects are atomic and sequentially consistent. The system guarantees that if multiple processes perform operations on the same shared object nearly simultaneously, the system chooses some order and all processes see the same order of events. Orca integrates shared data and synchronization in a way not present in pagebased DSM systems. Two kinds of synchronization are needed in parallel programs. The first kind is mutual exclusion synchronization, to keep two processes from executing the same critical region at the same time. In Orca, each operation on a shared object is effectively like a critical region because the system guarantees that the final result is the same as if all the critical regions were executed one at a time (i.e., sequentially). In this respect, an Orca object is like a distributed form of a monitor (Hoare, 1975). The other kind of synchronization is condition synchronization, in which a process blocks waiting for some condition to hold. In Orca, condition synchronization is done with guards. In the example of Fig. 8-46, a process trying to pop an item from an empty stack will be suspended until the stack is no longer empty. The Orca runtime system handles object replication, migration, consistency, and operation invocation. Each object can be in one of two states: single copy or replicated. An object in single-copy state exists on only one machine, so all requests for it are sent there. A replicated object is present on all machines containing a process using it, which makes read operations easier (since they can be done locally), at the expense of making updates more expensive. When an operation that modifies a replicated object is executed, it must first get a sequence number from a centralized process that issues them. Then a message is sent to each machine holding a copy of the object, telling it to execute the operation. Since all such updates bear sequence numbers, all machines just carry out the operations in sequence order, which guarantees sequential consistency. Globe Most DSM, Linda, and Orca systems run on local systems, that is within a single building or campus. However, it is also possible to build an applicationlevel shared-memory system on a multicomputer that runs worldwide. In the

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

643

Globe system, an object can be located in the address space of multiple processes at the same time, possibly on different continents (Kermarrec et al., 1998; Popescu et al., 2002; and Van Steen et al., 1999). To access a shared object’s data, user processes must go through its methods, which allows different objects to have different implementation strategies. For example, one option is to have a single copy of the data that is requested dynamically as needed (good for data frequently updated by a single owner). Another option is to have all the data located within each copy of the object and updates sent out to each copy by a reliable multicast protocol. What makes Globe somewhat ambitious is its goal to scale to a billion users and a trillion (possibly mobile) objects. Locating objects, managing them, and handling the scaling are crucial. Globe does this by having a general framework in which each object can nevertheless have its own replication strategy, security strategy, and so on. This avoids the one-size-fits-all problem present in other systems, while retaining the ease of programming offered by shared memory. Other wide-area distributed systems include Globus (Foster and Kesselman, 1998a; and Foster and Kesselman, 1998b) and Legion (Grimshaw and Wulf, 1996; and Grimshaw and Wulf, 1997), but these do not provide the illusion of shared memory as Globe does.

8.4.7 Performance The point of building a parallel computer is to make it go faster than a uniprocessor machine. If it does not achieve that simple goal, it is not worth having. Furthermore, it should achieve the goal in a cost-effective manner. A machine that is twice as fast as a uniprocessor at 50 times the cost is not likely to be a big seller. In this section we will examine some of the performance issues associated with parallel computer architectures. Hardware Metrics From a hardware perspective, the performance metrics of interest are the CPU and I/O speeds and the performance of the interconnection network. The CPU and I/O speeds are the same as in the uniprocessor case, so the key parameters of interest in a parallel system are those associated with the interconnect. There are two key items: latency and bandwidth, which we will now look at in turn. The roundtrip latency is the time it takes for a CPU to send a packet and get a reply. If the packet is sent to a memory, then the latency measures the time to read or write a word or block of words. If it is sent to another CPU, it measures the interprocessor communication time for packets of that size. Usually, the latency of interest is for minimal packets, often one word or a small cache line.

644

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

The latency is built up from several factors, and is different for circuitswitched, store-and-forward, virtual cut through, and wormhole-routed interconnects. For circuit switching, the latency is the sum of the setup time and the transmission time. To set up a circuit, a probe packet has to be sent out to reserve the resources and then report back. Once that has happened, the data packet has to be assembled. When it is ready, bits can flow at full speed, so if the total setup time is Ts , the packet size is p bits, and the bandwidth b bits/sec, the one-way latency is Ts + p/b. If the circuit is full duplex, then there is no setup time for the reply, so the minimum latency for sending a p bit packet and getting a p bit reply is Ts + 2p /b sec. For packet switching, it is not necessary to send a probe packet to the destination in advance, but there is still some internal setup time to assemble the packet, Ta . Here the one-way transmission time is Ta + p/b, but this is only the time to get the packet into the first switch. There is a finite delay within the switch, say Td and then the process is repeated to the next switch and so on. The Td delay is composed of both processing time and queueing delay, waiting for the output port to become free. If there are n switches, then the total one-way latency is given by the formula Ta + n(p /b + Td ) + p/b, where the final term is due to the copy from the last switch to the destination. The one-way latencies for virtual cut through and wormhole routing in the best case are close to Ta + p/b because there is no probe packet to set up a circuit, and no store-and-forward delay either. Basically, it is the initial setup time to assemble the packet, plus the time to push the bits out the door. In all cases, propagation delay has to be added, but that is usually small. The other hardware metric is bandwidth. Many parallel programs, especially in the natural sciences, move a lot of data around, so the number of bytes/sec that the system can move is critical to performance. Several metrics for bandwidth exist. We have seen one of them—bisection bandwidth—already. Another one is the aggregate bandwidth, which is computed by simply adding up the capacities of all the links. This number gives the maximum number of bits that can be in transit at once. Yet another important metric is the average bandwidth out of each CPU. If each CPU is capable of outputting 1 MB/sec, it does little good that the interconnect has a bisection bandwidth of 100 GB/sec. Communication will be limited by how much data each CPU can output. In practice, actually achieving anything even close to the theoretical bandwidth is very difficult. Many sources of overhead work to reduce the capacity. For example, there is always some per-packet overhead associated with each packet: assembling it, building its header, and getting it going. Sending 1024 4byte packets will never achieve the same bandwidth as sending one 4096-byte packet. Unfortunately, for achieving low latencies, using small packets is better, since large ones block the lines and switches too long. Thus there is an inherent conflict between achieving low average latencies and high-bandwidth utilization. For some applications, one is more important than the other and for other app-

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

645

lications it is the other way around. It is worth noting, however, that you can always buy more bandwidth (by putting in more or wider wires), but you cannot buy lower latencies. Thus it is generally better to err on the side of making latencies as short as possible, and worry about bandwidth later. Software Metrics Hardware metrics like latency and bandwidth look at what the hardware is capable of doing. However, users have a different perspective. They want to know how much faster their programs are going to run on a parallel computer than on a uniprocessor. For them, the key metric is speedup: how much faster a program runs on an n-processor system than on a one-processor system. Typically these results are shown in graphs like those of Fig. 8-47. Here we see several different parallel programs run on a multicomputer consisting of 64 Pentium Pro CPUs. Each curve shows the speedup of one program with k CPUs as a function of k. Perfect speedup is indicated by the dotted line, in which using k CPUs makes the program go k times faster, for any k. Few programs achieve perfect speedup, but some come close. The N-body problem parallelizes extremely well; awari (an African board game) does reasonably well; but inverting a certain skyline matrix does not go more than five times faster no matter how many CPUs are available. The programs and results are discussed in Bal et al. (1998). Part of the reason that perfect speedup is nearly impossible to achieve is that almost all programs have some sequential component, often the initialization phase, reading in the data, or collecting the results. Having many CPUs does not help here. Suppose that a program runs for T sec on a uniprocessor, with a fraction f of this time being sequential code and a fraction (1 − f ) being potentially parallelizable, as shown in Fig. 8-48(a). If the latter code can be run on n CPUs with no overhead, its execution time can be reduced from (1 − f )T to (1 − f )T/n at best, as shown in Fig. 8-48(b). This gives a total execution time for the sequential and parallel parts of fT + (1 − f )T/n. The speedup is just the execution time of the original program, T, divided by this new execution time: n Speedup = " """"""""" 1 + (n − 1)f For f = 0 we can get linear speedup, but for f > 0, perfect speedup is not possible due to the sequential component. This result is known as Amdahl’s law. Amdahl’s law is not the only reason perfect speedup is nearly impossible to achieve. Nonzero communication latencies, finite communication bandwidths, and algorithmic inefficiencies can also play a role. Also, even if 1000 CPUs were available, not all programs can be written to make use of so many CPUs, and the overhead in getting them all started may be significant. Furthermore, sometimes the best-known algorithm does not parallelize well, so a suboptimal algorithm must be used in the parallel case. This all said, for many applications, having the

646

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

60 N-body problem

50

Linear speedup

Speedup

40 Awari

30

20

10 Skyline matrix inversion 0

0

10

20

30 40 Number of CPUs

50

60

Figure 8-47. Real programs achieve less than the perfect speedup indicated by the dotted line.

n CPUs active

Potentially parallelizable part

f

1–f T (a)

…

Inherently sequential part

1 CPU active f

1–f

fT

(1 – f)T/n (b)

Figure 8-48. (a) A program has a sequential part and a parallelizable part. (b) Effect of running part of the program in parallel.

program run n times faster is highly desirable, even if it takes 2n CPUs to do it. CPUs are not that expensive, after all, and many companies live with considerably less than 100% efficiency in other parts of their businesses.

SEC. 8.4

647

MESSAGE-PASSING MULTICOMPUTERS

Achieving High Performance The most straightforward way to improve performance is to add more CPUs to the system. However, this addition must be done in such a way as to avoid creating any bottlenecks. A system in which one can add more CPUs and get correspondingly more computing power is said to be scalable. To see some of the implications of scalability, consider four CPUs connected by a bus, as illustrated in Fig. 8-49(a). Now imagine scaling the system to 16 CPUs by adding 12 more, as shown in Fig. 8-49(b). If the bandwidth of the bus is b MB/sec, then by quadrupling the number of CPUs, we have also reduced the available bandwidth per CPU from b/4 MB/sec to b/16 MB/sec. Such a system is not scalable. CPU

Bus (a)

(b)

(c)

(d)

Figure 8-49. (a) A 4-CPU bus-based system. (b) A 16-CPU bus-based system. (c) A 4-CPU grid-based system. (d) A 16-CPU grid-based system.

Now we do the same thing with a grid-based system, as shown in Fig. 8-49(c) and Fig. 8-49(d). With this topology, adding new CPUs also adds new links, so scaling the system up does not cause the aggregate bandwidth per CPU to drop, as it does with a bus. In fact, the ratio of links to CPUs increases from 1.0 with 4 CPUs (4 CPUs, 4 links) to 1.5 with 16 CPUs (16 CPUs, 24 links), so adding CPUs improves the aggregate bandwidth per CPU. Of course, bandwidth is not the only issue. Adding CPUs to the bus does not increase the diameter of the interconnection network or latency in the absence of traffic, whereas adding them to the grid does. For an n × n grid, the diameter is 2(n − 1), so the worst (and average) case latency increases roughly as the square root of the number of CPUs. For 400 CPUs, the diameter is 38, whereas for 1600 CPUs it is 78, so quadrupling the number of CPUs approximately doubles the diameter and thus the average latency. Ideally, a scalable system should maintain the same average bandwidth per CPU and a constant average latency as more and more CPUs are added. In practice, however, keeping enough bandwidth per CPU is doable, but in all practical designs, latency grows with size. Having it grow logarithmically, as in a hypercube, is about the best that can be done.

648

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

The problem with having latency grow as the system scales up is that latency is often fatal to performance in fine- and medium-grained applications. If a program needs data that are not in its local memory, there is often a substantial delay in getting them, and the bigger the system, the longer the delay, as we have just seen. This problem is equally true of multiprocessors as multicomputers, since in both cases the physical memory is invariably divided up into far-flung modules. As a consequence of this observation, system designers often go to great lengths to reduce, or at least hide, the latency, using several techniques we will now mention. The first latency-hiding technique is data replication. If copies of a block of data can be kept at multiple locations, accesses from those locations can be speeded up. One such replication technique is caching, in which one or more copies of data blocks are kept close to where they are being used, as well as where they ‘‘belong.’’ However, another strategy is to maintain multiple peer copies— copies that have equal status—as opposed to the asymmetric primary/secondary relationship used in caching. When multiple copies are maintained, in whatever form, key issues are where the data blocks are placed, when, and by whom. Answers range from dynamic placement on demand by the hardware, to intentional placement at load time following compiler directives. In all cases, managing consistency is an issue. A second technique for hiding latency is prefetching. If a data item can be fetched before it is needed, the fetching process can be overlapped with normal execution, so that when the item is needed, it will be there. Prefetching can be automatic or under program control. When a cache loads not only the word being referenced, but an entire cache line containing the word, it is gambling that the succeeding words are also likely to be needed soon. Prefetching can also be controlled explicitly. When the compiler realizes that it will need some data, it can put in an explicit instruction to go get them, and put that instruction sufficiently far in advance that the data will be there in time. This strategy requires that the compiler have a complete knowledge of the underlying machine and its timing, as well as control over where all data are placed. Such speculative LOAD instructions work best when it is known for sure that the data will be needed. Getting a page fault on a LOAD for a path that is ultimately not taken is very costly. A third technique that can hide latency is multithreading, as we have seen. If switching between processes can be made fast enough, for example, by giving each one its own memory map and hardware registers, then when one thread blocks waiting for remote data to arrive, the hardware can quickly switch to another one that is able to continue. In the limiting case, the CPU runs the first instruction from thread one, the second instruction from thread two, and so on. In this way, the CPU can be kept busy, even in the face of long memory latencies for the individual threads. A fourth technique for hiding latency is using nonblocking writes. Normally, when a STORE instruction is executed, the CPU waits until the STORE has

SEC. 8.4

MESSAGE-PASSING MULTICOMPUTERS

649

completed before continuing. With nonblocking writes, the memory operation is started, but the program just continues anyway. Continuing past a LOAD is harder, but with out-of-order execution, even that is possible.

8.5 GRID COMPUTING Many of today’s challenges in science, engineering, industry, the environment, and other areas, are large scale and interdisciplinary. Solving them requires the expertise, skills, knowledge, facilities, software, and data from multiple organizations, often in different countries. Some examples are as follows: 1. Scientists developing a mission to Mars. 2. A consortium building a complex product (e.g., a dam or aircraft). 3. An international relief team coordinating aid after a natural disaster. Some of these are long-term cooperations, others are more short term, but they all share the common thread of requiring separate organizations with their own resources and procedures to work together to achieve a common goal. Until recently, having different organizations, with different computers operating systems, data bases, and protocols work together to share resources and data has been very difficult. However, the growing need for large-scale interorganizational cooperation has led to the development of systems and technology for connecting widely-separated computers into what is called the grid. In a sense, the grid is the next step along the axis of Fig. 8-1. It can be thought of as a very large, international, loosely-coupled, heterogeneous, cluster. The goal of the grid is to provide a technical infrastructure to allow a group of organizations that share a common goal to form a virtual organization. This virtual organization has to be flexible, with a large and changing membership, permitting the members to work together in areas they deem appropriate, while allowing them to maintain control over their own resources to whatever degree they wish. To this end, grid researchers are developing services, tools, and protocols to enable these virtual organizations to function. The grid is inherently multilateral with many participants who are peers. It can be contrasted with existing computing frameworks. In the client-server model, a transaction involves two parties, the server, who offers some service, and the client, who wants to use the service. A typical example of the client-server model is the Web, in which users to go Web servers to find information. The grid also differs from peer-to-peer applications, in which pairs of individuals exchange files. E-mail is a common example of a peer-to-peer application. Because the grid is different from these models, it requires new protocols and technology. The grid needs access to a wide variety of resources. Each resource has a specific system and organization that owns it and which decides how much of the

650

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

resource to make available to the grid, during which hours, and to whom. In an abstract sense, what the grid is about is resource access and management. One way to model the grid is the layered hierarchy of Fig. 8-50. The fabric layer at the bottom is the set of components from which the grid is built. It includes CPUs, disks, networks, and sensors on the hardware side, and programs and data on the software side. These are the resources that the grid makes available in a controlled way. Layer

Function

Application

Applications that share managed resources in controlled ways

Collective

Discovery, brokering, monitoring and control of resource groups

Resource

Secure, managed access to individual resources

Fabric

Physical resources: computers, storage, networks, sensors, programs and data Figure 8-50. The grid layers.

One level higher, is the resource layer, which is concerned with managing the individual resources. In many cases, a resource participating in the grid has a local process that manages it and allows controlled access to it by remote users. This layer provides a uniform interface to higher layers for inquiring about the characteristics and status of individual resources, monitoring them, and using them in a secure way. Next is the collective layer, which handles groups of resources. One of its functions is resource discovery, by which a user can locate available CPU cycles, disk space, or specific data. The collective layer may maintain directories or other data bases to provide this information. It may also offer a brokering service by which the providers and users of services are matched up, possibly allocating scarce resources among competing users. The collective layer is also responsible for replicating data, managing the admission of new members and resources to the grid, accounting, and maintaining the policy data bases of who can use what. Still further up is the application layer, where the user applications reside. It uses the lower layers to acquire credentials proving its right to use certain resources, submit usage requests, monitor the progress of these requests, deal with failures, and notify the user of the results. Security is the key to a successful grid. Resource owners nearly always insist on maintaining tight control of their resources and want to determine who gets to use them, for how long, and how much. Without good security, no organization would make its resources available to the grid. On the other hand, if a user had to

SEC. 8.5

GRID COMPUTING

651

have a login account and password on every computer he wanted to use, using the grid would be unbearably cumbersome. Consequently, the grid has had to develop a security model to handle these concerns. One of the key characteristics of the security model is the single sign on. The first step in using the grid is to be authenticated and acquire a credential, a digitally signed document specifying on whose behalf the work is to be done. Credentials can be delegated, so that when a computation needs to create subcomputations, the child processes can also be identified. When a credential is presented at a remote machine, it has to be mapped onto the local security mechanism. On UNIX systems, for example, users are identified by 16-bit user IDs, but other systems have other schemes. Finally, the grid needs mechanisms to allow access policies to be stated, maintained, and updated. In order to provide interoperability between different organizations and machines, standards are needed, both in terms of the services offered and the protocols used to access them. The grid community has created an organization, the Global Grid Forum, to manage the standardization process. It has come up with a framework called OGSA (Open Grid Services Architecture) for positioning the various standards it is developing. Wherever possible, the standards utilize existing standards, for example, using WSDL (Web Services Definiton Language) for describing OGSA services. The services being standardized currently fall into eight broad categories as follows, but no doubt new ones will be created later. 1. Infrastructure services (enable communication between resources). 2. Resource management services (reserve and deploy resources). 3. Data services (move and replicate data to where it is needed). 4. Context services (describe required resources and usage policies). 5. Information services (get information about resource availability). 6. Self-Management services (support a stated quality of service). 7. Security services (enforce security policies). 8. Execution management services (manage workflow). There is much more that could be said about the grid, but space limitations prevent us from pursuing this topic further. For more information about the grid, see (Berman et al., 2003; Foster and Kesselman, 2003; and Foster et al., 2002).

8.6 SUMMARY It is getting increasingly difficult to make computers go faster by just revving up the clock due to increased heat dissipation problems and other factors. Instead, designers are looking to parallelism for speedup. Parallelism can be introduced at

652

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

many different levels, from very low, where the processing elements are very tightly coupled, to very high, where they are very loosely coupled. At the bottom level is on-chip parallelism, in which parallel activities occur on a single chip. One form of on-chip parallelism is instruction-level parallelism, in which one instruction or a sequence of instructions issue multiple operations that can be executed in parallel by different functional units. A second form of on-chip parallelism is multithreading, in which the CPU can switch back and forth among multiple threads on an instruction by instruction basis, creating a virtual multiprocessor. A third form of on-chip parallelism is the single-chip multiprocessor, in which two or more cores are placed on the same chip to allow them to run at the same time. One level up we find the coprocessors, typically plug-in boards that add extra processing power in some specialized area such as network protocol processing or multimedia. These extra processors relieve the main CPU of work, allowing it to do other things while they are performing their specialized tasks. At the next level, we find the shared-memory multiprocessors. The systems contain two or more full-blown CPUs that share a common memory. UMA multiprocessors communicate via a shared (snooping) bus, a crossbar switch, or a multistage switching network. They are characterized by having a uniform access time to all memory locations. In contrast, NUMA multiprocessors also present all processes with the same shared address space, but here remote accesses take appreciably longer than local ones. Finally, COMA multiprocessors are yet another variation, in which cache lines move around the machine on demand, but have no real home as in the other designs. Multicomputers are systems with many CPUs that do not share a common memory. Each one has its own private memory, with communication by message passing. MPPs are large multicomputers with specialized communication networks such as IBM’s BlueGene/L. Clusters are simpler systems using off-theself-components, such as the engine that powers Google. Multicomputers are often programmed using a message-passing package such as MPI. An alternative approach is to use application-level shared memory such as a page-based DSM system, the Linda tuple space, or Orca or Globe objects. DSM simulates shared memory at the page level, making it similar to a NUMA machine, except with a much greater penalty for remote references. Finally, at the highest level, and the most loosely coupled, are the grids. These are systems in which entire organizations are hooked together over the Internet to share compute power, data, and other resources. PROBLEMS 1. Pentium instructions can be as long as 17 bytes. Is the Pentium a VLIW CPU? 2. What are the clipped values of 96, −9, 300, and 256 when the clipping range is 0–255?

CHAP. 8

PROBLEMS

653

3. Are the following TriMedia instructions allowed, and if not, why not? (a) Integer add, integer subtract, load, floating add, load immediate (b) Integer subtract, integer multiply, load immediate, shift, shift (c) Load immediate, floating add, floating multiply, branch, load immediate 4. Figure 8-7(d) and (e) show 12 cycles of instructions. For each one, tell what happens in the following three cycles. 5. On a particular CPU, an instruction that misses the level 1 cache but hits the level 2 cache takes k cycles in total. If multithreading is used to mask level 1 cache misses, how many threads must be run at once using fine-grained multithreading to avoid dead cycles? 6. One morning, the queen bee of a certain beehive calls in all her worker bees and tells them that today’s assignment is to collect marigold nectar. The workers then fly off in different directions looking for marigolds. Is this an SIMD or an MIMD system? 7. During our discussion of memory consistency models, we said that a consistency model is a kind of contract between the software and the memory. Why is such a contract needed? 8. Consider a multiprocessor using a shared bus. What happens if two processors try to access the global memory at exactly the same instant? 9. Suppose that for technical reasons it is only possible for a snooping cache to snoop on address lines, not data lines. Would this change affect the write through protocol? 10. As a simple model of a bus-based multiprocessor system without caching, suppose that one instruction in every four references memory, and that a memory reference occupies the bus for an entire instruction time. If the bus is busy, the requesting CPU is put into a FIFO queue. How much faster will a 64-CPU system run than a 1-CPU system? 11. The MESI cache coherence protocol has four states. Other write-back cache coherence protocols have only three states. Which of the four MESI states could be sacrificed, and what would the consequences of each choice be? If you had to pick only three states, which would you pick? 12. Are there any situations with the MESI cache coherence protocol in which a cache line is present in the local cache but for which a bus transaction is nevertheless needed? If so, explain. 13. Suppose that there are n CPUs on a common bus. The probability that any CPU tries to use the bus in a given cycle is p. What is the chance that a. The bus is idle (0 requests). b. Exactly one request is made. c. More than one request is made. 14. How many crossbar switches does a full Sun Fire E25K have? 15. Suppose that the wire between switch 2A and switch 3B in the omega network breaks. Who is cut off from whom?

654

PARALLEL COMPUTER ARCHITECTURES

CHAP. 8

16. Hot spots (heavily referenced memory locations) are clearly a major problem in multistage switching networks. Are they also a problem in bus-based systems? 17. An omega switching network connects 4096 RISC CPUs, each with a 60-nsec cycle time, to 4096 infinitely fast memory modules. The switching elements each have a 5nsec delay. How many delay slots are needed by a LOAD instruction? 18. Consider a machine using an omega switching network, like the one shown in Fig. 829. Suppose that the program and stack for processor i is kept in memory module i. Propose a slight change in the topology that makes a large difference in the performance (the IBM RP3 and BBN Butterfly use this modified topology). What disadvantage does your new topology have compared to the original? 19. In a NUMA multiprocessor, local memory references take 20 nsec and remote references take 120 nsec. A certain program makes a total of N memory references during its execution, of which 1 percent are to a page P. That page is initially remote, and it takes C nsec to copy it locally. Under what conditions should the page be copied locally in the absence of significant use by other processors? 20. Consider a CC-NUMA multiprocessor like that of Fig. 8-31 except with 512 nodes of 8 MB each. If the cache lines are 64 bytes, what is the percentage overhead for the directories? Does increasing the number of nodes increase the overhead, decrease the overhead, or leave it unchanged? 21. For each topology shown in Fig. 8-35, compute the diameter of the network. 22. For each topology shown in Fig. 8-35, determine the degree of fault tolerance each one has, defined as the maximum number of links that can be lost without partitioning the network in two. 23. Consider the double torus topology of Fig. 8-35(f) but expanded to a size of k × k. What is the diameter of the network? Hint: Consider odd k and even k separately. 24. An interconnection network is in the form of an 8 × 8 × 8 cube. Each link has a fullduplex bandwidth of 1 GB/sec. What is the bisection bandwidth of the network? 25. Amdahl’s law limits the potential speedup achievable on a parallel computer. Compute, as a function of f, the maximum possible speedup as the number of CPUs approaches infinity. What are the implications of this limit for f = 0.1? 26. Figure 8-49 shows how scaling fails with a bus but succeeds with a grid. Assuming that each bus or link has a bandwidth b, compute the average bandwidth per CPU for each of the four cases. Then scale each system to 64 CPUs and repeat the calculations. What is the limit as the number of CPUs goes to infinity? 27. In the text, three variations of send were discussed: synchronous, blocking, and nonblocking. Give a fourth method that is similar to blocking send, but has slightly different properties. Give an advantage and a disadvantage of your method as compared to blocking send. 28. Consider a multicomputer running on a network with hardware broadcasting, such as Ethernet. Why does the ratio of read operations (those not updating internal state variables) to write operations (those updating internal state variables) matter?

9 READING LIST AND BIBLIOGRAPHY

In the preceding eight chapters, a large number of topics were discussed in various degrees of detail. This chapter is meant to assist readers interested in pursuing further their study of computer organization. Section 9.1 contains a list of suggested readings arranged by chapter. Section 9.2 is an alphabetical bibliography of all books and articles cited in this book.

9.1 SUGGESTIONS FOR FURTHER READING Below are some suggested readings keyed by chapter. Most of them are textbooks, tutorials, or survey-type articles.

9.1.1 Introduction and General Works Borkar, ‘‘Getting Gigascale Chips’’ Moore’s law will probably hold for at least another decade, potentially leading to billion-transistor chips. Such chips present challenges and opportunities. In this article, one of Intel’s leading researchers discusses future challenges, such as energy dissipation, higher resistance and capacitance from the ever smaller and closer wires, and so on. He thinks the future lies in multithreading, multiprocessors on a chip, and better memory structures, rather than purely faster clock speeds. 655

656

READING LIST AND BIBLIOGRAPHY

CHAP. 9

Colwell, The Pentium Chronicles Robert Colwell was the leader of the team that designed the Pentium. In this book, he talks about the people, the passion and the politics behind this chip. Hamacher et al., Computer Organization, 5th ed. A traditional textbook on computer organization the CPU, memory, I/O, arithmetic, and peripherals. The main examples are the 68000 and the PowerPC. Heath, Embedded Systems Design Practically everything that runs on electricity and costs more than $50 has a computer in it nowadays. These embedded systems are the subject of this book. Starting with the basics of embedded processors, memory, and peripherals, it then moves on to interfacing, real-time operating systems, software, and debugging. Hennessy and Patterson, Computer Architecture A Quantitative Approach, 3rd ed. This massive graduate textbook goes into great detail on how to design a processor and its memory. The emphasis is on achieving high performance, especially by exploiting parallelism and pipelining. If you want to know everything about designing a high-end CPU, this book is the place to look. Null and Lobur, The Essentials of Computer Organization and Architecture Another textbook on Computer Organization covering many of the same topics as this book, but with less depth. Patterson and Hennessy, Computer Organization and Design, 3rd ed. No longer 1000 pages like the 2nd edition, the 3rd edition has transferred a great deal of the text to the accompanying CD-ROM. The material remaining in the book covers many aspects of computer architecture, including arithmetic, performance, the datapath, pipelining, memorie, peripherals, and clusters. Although the Pentium 4 is discussed in several places, the main example used to explain concepts is the MIPS processor, which was designed by Hennessy and was the first commercial RISC machine, introduced in 1985. Price, ‘‘A History of Calculating Machines’’ Although modern computers started with Babbage in the 19th century, human beings have been computing since the dawn of civilization. This fascinating illustrated article traces the history of counting, mathematics, calendars, and computation from 3000 B.C. to the start of the 20th century. Slater, Portraits in Silicon Why didn’t Dennis Ritchie turn in his Ph.D. thesis at Harvard? Why did Steve Jobs become a vegetarian? The answers lie in this fascinating book that contains short biographies of 34 people who shaped the computer industry, from Charles Babbage to Donald Knuth.

SEC. 9.1

SUGGESTIONS FOR FURTHER READING

657

Stallings, Computer Organization and Architecture, 6th ed. A general text on computer architecture. Some of the topics treated in this book are also covered in Stallings’ book. Wilkes, ‘‘Computers Then and Now’’ A personal history of computers from 1946 to 1968 by pioneer computer designer and inventor of microprogramming, Maurice Wilkes. He tells of the early battles between the ‘‘space cadets,’’ who believed in automatic programming (pre-FORTRAN compilers), and the traditionalists, who preferred to do their programming in octal.

9.1.2 Computer Systems Organization Buchanan and Wilson, Advanced PC Architecture Although somewhat disorderly, this book does cover a lot of ground concerning processors, buses (PCI, SCSI, and USB), ports (game, parallel, and serial) and other PC components. Ng, ‘‘Advances in Disk Technology: Performance Issues’’ People have been predicting the end of the magnetic disk for at least 20 years. So far they still seem to be with us. And according to this paper, their technology is advancing rapidly, so they are likely to be around for years to come. Messmer, The Indispensable PC Hardware Book, 4th ed. At 1296 pages (divided into 37 chapters and 7 appendices), this book may or may not be indispensable, but it sure is thick. Just about everything there is to know about 80x86 processors, memories, buses, support chips, and peripherals is here in gory detail. If you have read and digested Norton and Goodman’s book (below) and want to move on to the next level of technical detail, start here. Norton and Goodman, Inside the PC, 8th ed. Most books on PC hardware are written for people with an EE degree and are difficult to read for software-oriented people. This one is different. It explains PC hardware in a technical, but highly accessible way. Topics include the CPU, memory, buses, disks, displays, I/O devices, mobile PCs, networking and more. A rare and worthwhile book. Robinson, ‘‘Toward the Age of Smarter Storage’’ Storage has come a long way since the days of core memories and punched cards. This short article looks at where storage technology has been, where it is now, and where it is going. Scheible, ‘‘A Survey of Storage Options’’ Another rundown of memory technology, only focusing in what is available now. It discusses the various flavors of RAM, flash memory, tape, hard disks, floppy disks, CDs, adb DVDs.

658

READING LIST AND BIBLIOGRAPHY

CHAP. 9

Stan and Skadron, ‘‘Power-Aware Computing’’ As computers are becoming too powerful—literally. They are consuming too much power, which is a growing problem in an increasingly mobile world. This paper is the guest editors’ introduction to a special issue of IEEE Computer Magazine on power-aware computing. Triebel, The 80386, 80486, and Pentium Processor It is a bit hard to classify this book as it deals with hardware, software, and interfacing. Since the author is with Intel, let us call it a hardware book. It tells all about the processors, memories, I/O devices, and interfacing of the 80x86 chips, but also how to program them in assembly language. Although it has a mere 915 pages, it contains almost as much material as Messmer’s book because the pages are bigger.

9.1.3 The Digital Logic Level Floyd, Digital Fundamentals, 8th ed. For hardware-oriented readers who want to pursue their study of the digital logic level in more detail, this huge, lavishly illustrated four-color book is a real gem. Chapters cover combinational logic, programmable logic devices, flip flops, shift registers, memories, interfacing, and much more. Mano and Kime, Logic and Computer Design Fundamentals, 3rd ed. Although this book does not have the elegance layout of Floyd’s book, it is also a good reference for the digital logic level. It covers combinational and sequential circuits, registers, memories, CPU design and I/O. Mayhew and Krishnan, ‘‘PCI Express and Advanced Switching’’ PCI Express is likely to replace the PCI bus in the near future and this paper gives a tutorial on it covering the layers, flow control, virtual channels, switching, and routing. Mazidi and Mazidi, The 80x86 IBM PC and Compatible Computers, 4th ed. For readers interested in understanding all the chips inside a PC, this book has entire chapters on the major chips, as well as a wealth of other information about the IBM PC hardware and assembly language programming. Roth, Fundamentals of Logic Design This textbook covers the basics of digital logic design, from Boolean algebra through gates, counters, adders, flip-flops, and other kinds of combinational and sequential circuits.

SEC. 9.1

SUGGESTIONS FOR FURTHER READING

659

9.1.4 The Microarchitecture Level Burger and Goodman, ‘‘Billion-Transistor Architectures: There and Back Again’’ Suppose that in 1997 somebody gave you a billion transistors and said: ‘‘Design a chip.’’ What kind of microarchitecture would you design. In Sept. 1997, seven top researchers in architecture who were asked this question had their views published in IEEE Computer Magazine. Seven years later their predictions are compared to the current state of the art. Handy, The Cache Memory Book, 2nd ed. Cache design is sufficiently important that entire books on the subject now exist. This one discusses logical versus physical caches, line size, write-through versus write-back policies, unified versus split caches, as well as software issues. It also has a chapter on multiprocessor cache coherence. Johnson, Superscalar Microprocessor Design For readers interested in the details of superscalar CPU design, this book is the place to start. It covers instruction fetching and decoding, out of order instruction issue, register renaming, reservation stations, branch prediction, and more. Shriver and Smith, ‘‘The Anatomy of a High-Performance Microprocessor’’ For a detailed study of a modern CPU chip at the microarchitecture level, this book is a good bet. It examines the AMD K6 chip, a Pentium clone, in detail, emphasizing the pipeline, instruction scheduling, and performance optimiations. Sima, ‘‘Superscalar Instruction Issue’’ Superscalar instruction issue is increasingly important in modern CPUs. We touched upon some of the issues in this chapter, such as renaming and speculative execution. In this paper, these and a number of other issues are examined. Wilson, ‘‘Challenges and Trends in Processor Design’’ Is processor design dead in the water? No way. Six top CPU architects from Sun, Cyrix, Motorola, Mips, Intel, and Digital tell where they think CPU design is going in the next few years. This will be fun reading in 2008 (but it is also worthwhile now).

9.1.5 The Instruction Set Architecture Level Antonakos, The Pentium Microprocessor The first nine chapters of this book deal with how to program the Pentium in assembly language. The last two deal with the Pentium hardware. Numerous code fragments are given and the BIOS is treated throughout.

660

READING LIST AND BIBLIOGRAPHY

CHAP. 9

Ayala, The 8051 Microcontroller, 3rd ed. If you are interested in learning to program the 8051, this book is a reasonable place to start. Bryant and O’Hallaron, Computer Systems A Programmer’s Perspective Although somewhat disorganized, this book covers a lot of ground concerning the ISA level, including arithmetic, different kinds of instructions, control flow, and program optimization. Paul, SPARC Architecture, Assembly Language, Programming, and C Wonder of wonders, here is a book about assembly language programming that is not about the Intel 80x86 line. Instead it is about the SPARC and how to program it. Weaver and Germond, The SPARC Architecture Manual With the growing internationalization of the computer industry, standards are becoming more important all the time, so it is important to be familiar with them. This one is the Version 9 SPARC definition and gives a good idea of what a standard looks like and is also very informative about how 64-bit SPARCs work.

9.1.6 The Operating System Machine Level Hart, ‘‘Win32 System Programming’’ Unlike nearly all other books on Windows, this one does not dwell on (or even cover) the graphical user interface. Instead it focuses on the systems calls offered by Windows and how to use them for file access, memory management, process management, interprocess communication, threads, I/O, and other topics. Jacob and Mudge, ‘‘Virtual Memory: Issues of Implementation’’ For a good and modern introduction on virtual memory, look here. It explains various page table and TLB structures and illustrates the ideas using the MIPS, PowerPC, and Pentium processors. McKusick et al., Design and Implementation of the 4.4 BSD Operating System Unlike most books on UNIX, this one starts out with a photo of the four authors at a USENIX Conference, three of whom wrote much of 4.4 BSD and are eminently qualified to explain its inner workings. The book covers the system calls, processes, I/O, and has an especially good section on networking. Ritchie and Thompson, ‘‘The UNIX Time-Sharing System’’ This is the original published paper on UNIX. It is still well worth reading. From this small seed has grown a great operating system.

SEC. 9.1

SUGGESTIONS FOR FURTHER READING

661

Russinovich and Solomon, Inside Microsoft Windows, 4th ed. If you want to know how Windows works internally, this is your best bet. It discusses the system architecture, system mechanisms, processes, threads, memory management, security, I/O, cache, and file system, among other topics. The book is intended as an in-depth study for CS students or IT professionals. Tanenbaum and Woodhull, Operating Systems: Design and Implementation, 2nd ed Unlike most books on operating systems that only deal with the theory, this one covers all the relevant theory and illustrates it by discussing the actual code of a UNIX-like operating system, MINIX, that runs on the IBM PC and other computers. The heavily annotated source code is listed in an appendix.

9.1.7 The Assembly Language Level Levine, Linkers and Loaders If your thing is linkers and loaders, and you revel in the various object formats around, the difference between static and dynamic linking, and the various library formats, this book is for you. Saloman, Assemblers and Loaders Everything you might want to know about how one-pass and two-pass assemblers work, as well as how linkers and loaders work is here. Macros and conditional assembly are also covered.

9.1.8 Parallel Computer Architectures Adve and Gharachorloo, ‘‘Shared Memory Consistency Models: A Tutorial’’ Many modern computers, especially multiprocessors, support a memory model weaker than sequential consistency. This tutorial discusses various models and explains how they work. It also states and rebuts numerous myths about weakly consistent memory. Comer, Network Systems Design The first part of this book is about traditional packet processing in networks, but the second part introduces network processors and describes their purpose, architectures, and design trade-offs. The third part looks at the Agere network processor as a case study. Dally and Towles, Principles and Practices of Interconnection Networks If you are interested in interconnection networks, this is the place to look. After an introduction to topology, butterfly networks, torus networks, and nonblocking networks are covered. Then come quite a few chapters on routing, flow control, buffering, deadlock, and related issues.

662

READING LIST AND BIBLIOGRAPHY

CHAP. 9

Dongarra et al., The Sourcebook of Parallel Computing Programming multiprocessors and clusters is appreciably different from programming uniprocessors. In this book, seven leading experts in the field of parallel programming cover various aspects of parallel programming, including parallel architectures, software technologies, parallel algorithms, and some applications. Hill, ‘‘Multiprocessors Should Support Simple Memory-Consistency Models’’ Relaxed memory semantics is a hot and controversial topic in multiprocessor memory design. Weaker models allow certain hardware optimizations, like making memory references out of order, but make programming harder. In this paper, the author discusses many issues relevant to memory consistency and then concludes that relaxed memory is more trouble than it is worth. Hwang and Xu, Scalable Parallel Computing By treating both the hardware and the software, the authors manage to give a comprehensive, yet readable account of parallel computing. Topics include UMA and NUMA multiprocessors, MPPs and COWs, message-passing, and data parallel programming. Lawton, ‘‘Will Network Processor Units Live up to Their Promise?’’ While network processors promise to speed up packet processing, their success is not guaranteed. In this paper, the author looks at the technology and some of the factors that may determine its success or failure. McKnight et al., ‘‘Wireless Grids’’ Grids have barely gotten started, yet the next generation—wireless grids—is already on the horizon. Like regular grids, these aim at sharing resources across organizations to create virtual organizations, only they use wireless technology to make the resources available to mobile users. The two articles following this one are also on wireless grids. Pfister, In Search of Clusters, 2nd ed. Although the definition of a cluster does not occur until page 72 (a bunch of whole computers working together), it apparently includes all the usual multiprocessor and multicomputer systems. Their hardware, software, performance, and availability are all studied in detail. The reader should be warned however, that while the author’s cutesy-poo writing style is amusing at first, by page 500 the novelty wears off. Snir et al., MPI: The Complete Reference Manual The title says it all. If you want to learn to program in MPI, look here. The book covers point-to-point and collective communication, communicators, environmental management, profiling, and more.

SEC. 9.1

SUGGESTIONS FOR FURTHER READING

663

Stenstrom et al., ‘‘Trends in Shared Memory Multiprocessing’’ Although shared-memory multiprocessors are often thought as supercomputers for massive scientific calculations, in reality that is only a tiny fraction of their market. In this paper, the authors discuss where the market for these machines really lies, and what implications that has for their architecture. Ungerer et al., ‘‘A Survey of Processors with Explicit Multithreading’’ For each of the major kinds of multithreading—fine-grained, coarse-grained, and simultaneous—the paper explains how it works and gives numerous examples of academic and commercial computers that use the technique. Wolf, ‘‘The Future of Multiprocessor Systems-on-Chips,’’ After presenting three current system-on-a-chip designs, the author goes on to look at hardware and software challenges for future systems. Hardware problems include real-time and energy dissipation concerns. Software concerns include operating system issues and potential security problems.

9.1.9 Binary and Floating-Point Numbers Cody, ‘‘Analysis of Proposals for the Floating-Point Standard’’ Some years ago, IEEE designed a floating-point architecture that has become the de facto standard for all modern CPU chips. Cody discusses the various issues, proposals, and controversies that came up during the standardization process. Koren ‘‘Computer Arithmetic Algorithms’’ A complete book on arithmetic, with an emphasis on algorithms for fast addition, multiplication and division. Strongly recommended for anyone who thinks he learned everything there is to know about arithmetic in sixth grade. IEEE, Proc. of the n-th Symposium on Computer Arithmetic Contrary to popular opinion, arithmetic is an active research area with many scientific papers written by and for arithmetic specialists. In this symposium series, advances in high-speed addition and multiplication, VLSI arithmetic hardware, coprocessors, fault tolerance, and rounding, among other topics, are presented. Knuth, Seminumerical Algorithms, 3rd ed. A wealth of material about positional number systems, floating-point arithmetic, multiple-precision arithmetic, and random numbers. This material requires and deserves careful study.

664

READING LIST AND BIBLIOGRAPHY

CHAP. 9

Wilson, ‘‘Floating-Point Survival Kit’’ A nice introduction to floating-point numbers and standards for people who think that the world ends at 65,535. Some popular floating-point benchmarks, such as Linpack, are also discussed.

9.1.10 Assembly Language Programming Blum, Professional Assembly Language A guide to programming the Pentium in assembly language for professionals. Since the book is for professionals, he assumes you are running Linux on your Pentium and focuses on the Linux assembler and GNU tools as well as discussing the Linux system calls. Irvine, Assembly Language for Intel-Based Computers, 4th ed. Programming the Intel CPUs in assembly language is the subject of this book. Also covered are I/O programming, macros, files, linking, interrupts, and many other related topics.

9.2 ALPHABETICAL BIBLIOGRAPHY ADAMS, M., and DULCHINOS, D.: ‘‘OpenCable,’’ IEEE Commun. Magazine, vol. 39, pp.

98-105, June 2001.

ADIGA, N.R. et al.: ‘‘An Overview of the BlueGene/L Supercomputer,’’ Proc. Supercom-

puting 2002, ACM, pp. 1-22, 2002.

ADVE, S.V., and CHARACHORLOO, K.: ‘‘Shared Memory Consistency Models: A

Tutorial,’’ IEEE Computer Magazine, vol. 29, pp. 66-76, Dec. 1996.

ADVE, S.V., and HILL, M.: ‘‘Weak Ordering: A New Definition,’’ Proc. 17th Ann. Int’l

Symp. on Computer Arch., ACM, pp. 2-14, 1990.

AGERWALA, T., and COCKE, J.: ‘‘High Performance Reduced Instruction Set Proces-

sors,’’ IBM T.J. Watson Research Center Technical Report RC12434, 1987.

ALAMELDEEN, A.R., and Wood, D.A.: ‘‘Adaptive Cache Compression for High-

Performance Processors,’’ Proc. 31st Ann. Int’l Sym. on Computer Arch. ACM, pp. 212-223, 2004.

ALMASI, G.S. et al.: ‘‘System Management in the BlueGene/L Supercomputer,’’ Proc.

17th Int’l Parallel and Distr. Proc. Symp., IEEE, 2003a.

ALMASI, G.S. et al.: ‘‘An Overview Of The Bluegene/L System Software Organization,’’

Par. Proc. Letters, vol. 13, 561-574, April 2003b.

SEC. 9.2

ALPHABETICAL BIBLIOGRAPHY

665

AMZA, C., COX, A., DWARKADAS, S., KELEHER, P., LU, H., RAJAMONY, R., YU, W., ZWAENEPOEL, W.: ‘‘TreadMarks: Shared Memory Computing on a Network of

Workstations,’’ IEEE Computer Magazine, vol. 29, pp. 18-28, Feb. 1996.

ANDERSON, D.: Universal Serial Bus System Architecture, Reading, MA: Addison-

Wesley, 1997.

ANDERSON, D., BUDRUK, R., and SHANLEY, T.: PCI Express System Architecture,

Reading, MA: Addison-Wesley, 2004.

ANDERSON, T.E., CULLER, D.E., PATTERSON, D.A., and the NOW team: ‘‘A Case for

NOW (Networks of Workstations),’’ IEEE Micro Magazine, vol. 15, pp. 54-64, Jan. 1995.

ANTONAKOS, J.L.: The Pentium Microprocessor, Upper Saddle River, NJ: Prentice Hall,

1997.

AUGUST, D.I., CONNORS, D.A., MSHLKE, S.A., SIAS, J.W., CROZIER, K.M., CHENG, B.-C., EATON, P.R., OLANIRAN, Q.B., and HWU, W.-M.: ‘‘Integrated Predicated and

Speculative Execution in the IMPACT EPIC Architecture,’’ Proc. 25th Ann. Int’l Symp. on Computer Arch., ACM, pp. 227-237, 1998.

AYALA, K: The 8051 Microcontroller, 3rd ed., Clifton Park, NY: Thomson Delmar

Learning, 2004.

BAL, H.E.: Programming Distributed Systems, Hemel Hempstead, England: Prentice Hall

Int’l, 1991.

BAL, H.E., BHOEDJANG, R., HOFMAN, R, JACOBS, C., LANGENDOEN, K., RUHL, T., and KAASHOEK, M.F.: ‘‘Performance Evaluation of the Orca Shared Object Sys-

tem,’’ ACM Trans. on Computer Systems, vol. 16, pp. 1-40, Jan.-Feb. 1998.

BAL, H.E., KAASHOEK, M.F., and TANENBAUM, A.S.: ‘‘Orca: A Language for Parallel

Programming of Distributed Systems,’’ IEEE Trans. on Software Engineering, vol. 18, pp. 190-205, March 1992.

BAL, H.E., and TANENBAUM, A.S.: ‘‘Distributed Programming with Shared Data,’’ Proc.

1988 Int’l Conf. on Computer Languages, IEEE, pp. 82-91, 1988.

BARROSO, L.A., DEAN, J., HOLZLE, U.: ‘‘Web Search for a Planet: The Google Cluster

Architecture,’’ IEEE Micro Magazine, vol. 23, pp. 22-28, March-April 2003.

BECHINI, A., CONTE, T.M., and PRETE, C.A.: ‘‘Opportunities and Challenges in Embed-

ded Systems,’’ IEEE Micro Magazine, vol. 24, pp. 8-9, July-Aug. 2004.

BENINI, L., and DE MICHELI, G.: ‘‘Networks on Chips" A New SoC Paradigm,’’ IEEE

Computer Magazine, vol. 35, pp. 70-78, Jan. 2002.

BERMAN, F., FOX, G., and HEY, A.J.G.: ‘‘Grid Computing: Making the Global Infrast-

tructure a Reality,’’ Hoboken, NJ: John Wiley, 2003.

BJORNSON, R.D.: ‘‘Linda on Distributed Memory Multiprocessors,’’ Ph.D. Thesis, Yale

Univ., 1993.

666

READING LIST AND BIBLIOGRAPHY

CHAP. 9

BLUM, R.: Professional Assembly Language, New York: Wiley, 2005. BLUMRICH, M., CHEN, D., CHIU, G., COTEUS, P., GARA, A., GIAMPAPA, M.E., HARING, R.A., HEIDELBERGER, P., HOENICKE, D., KOPCSAY, G.V., OHMACHT, M., STEINMACHER-BUROW, B.D., TAKKEN, T., VRANSAS, P., and LIEBSCH, T.: ‘‘An

Overview of the BlueGene/L System,’’, IBM J. Research and Devel., vol. 49, MarchMay, 2005.

BORKAR, S.: ‘‘Getting Gigascale Chips,’’ Queue, pp. 26-33, Oct. 2003. BOSE, P.: ‘‘Computer architecture research: Shifting priorities and newer challenges,’’

IEEE Micro Magazine, vol. 24, p. 5, Nov.-Dec. 2004.

BOUKNIGHT, W.J., DENENBERG, S.A., MCINTYRE, D.E., RANDALL, J.M., SAMEH, A.H., and SLOTNICK, D.L.: ‘‘The Illiac IV System,’’ Proc. IEEE, pp. 369-388, April

1972.

BRIGHTWELL, R., CAMP, W., COLE, B., DEBENEDICTIS, LELAND, R, and TOMPKINS, J.: ‘‘Architectural Specification for Massively Parallel Computers—An Experience

and Measurement-Based Approach,’’ [To be published].

BRYANT, R.E., and O’HALLARON, D.: Computer Systems A Programmer’s Perspective

Upper Saddle River, NJ: Prentice Hall, 2003.

BUCHANAN, W., WILSON, A.: Advanced PC Architecture, Reading, MA: Addison-

Wesley, 2001.

BURGER, D., and GOODMAN, J.R.: ‘‘Billion-Transistor Architectures: There and Back

Again,’’ IEEE Computer Magazine, vol. 37, pp. 22-28, March 2004.

BURKHARDT, H., FRANK, S., KNOBE, B., and ROTHNIE, J.: ‘‘Overview of the KSR-1

Computer System,’’ Technical Report KSR-TR-9202001, Kendall Square Research Corp, Cambridge, MA, 1992.

CAIN, H., and LIPASTI, M.: ‘‘Memory Ordering: A Value-Based Approach,’’ Proc. 31th Ann. Int’l Symp. on Computer Arch., ACM, pp. 90-101, 2004. CALCUTT, D., COWAN, F., and PARCHIZADEH, H.: 8051 Microcontrollers : An Appli-

cations Based Introduction, Oxford: Newnes, 2004.

CARRIERO, N., and GELERNTER, D.: ‘‘Linda in Context,’’ Commun. of the ACM, vol.

32, pp. 444-458, April 1989.

CHARLESWORTH, A.: ‘‘The Sun Fireplane Interconnect,’’ IEEE Micro Magazine, vol.

22, pp. 36-45, Jan.-Feb. 2002.

CHARLESWORTH, A.: ‘‘The Sun Fireplane Interconnect,’’ Proc. Conf. on High Perf. Net-

working and Computing, ACM, 2001.

CHARLESWORTH, A., PHELPS, A., WILLIAMS, R., and GILBERT, G.: ‘‘Gigaplane-XB:

Extending the Ultra Enterprise Family,’’ Proc. Hot Interconnects V, IEEE, 1998.

SEC. 9.2

ALPHABETICAL BIBLIOGRAPHY

667

CHEN, L., DROPSHO, S., ALBONESI, D.H.: ‘‘Dynamic Data Dependence Tracking and its

Application to Branch Prediction,’’ Proc. Ninth Int’l Symp. on High-Performance Computer Arch., IEEE, pp. 65-78, 2003.

CHOU, Y., FAHS, B., ABRAHAM, S.: ‘‘ Microarchitecture Optimizations for Exploiting

Memory-Level Parallelism,|*(CQ Proc. 31st Ann. Int’l Symp. on Computer Arch., ACM, pp. 76-77, 2004.

CLAASEN, T.A.C.M.: ‘‘System on a Chip: Changing IC Design Today and in the Future,’’

IEEE Micro Magazine, vol. 23, pp. 20-26, May-June 2003.

CODY, W.J.: ‘‘Analysis of Proposals for the Floating-Point Standard,’’ IEEE Computer

Magazine, vol. 14, pp. 63-68, March 1981.

COHEN, D.: ‘‘On Holy Wars and a Plea for Peace,’’ IEEE Computer Magazine, vol. 14,

pp. 48-54, Oct. 1981.

COLWELL, R: The Pentium Chronicles New York: WIley, 2005. COMER, D.E.: ‘‘Network Systems Design Using Network Processors: Agere Version,’’

Upper Saddle River, NJ: Prentice Hall, 2005.

CORBATO ´ , F.J.: ‘‘PL/1 as a Tool for System Programming,’’ Datamation, vol. 15, pp.

68-76, May 1969.

CORBATO ´ , F.J., and VYSSOTSKY, V.A.: ‘‘Introduction and Overview of the MULTICS

System,’’ Proc. FJCC, pp. 185-196, 1965.

CROWLEY, P., FRANKLIN, M.A., HADIMIOGLU, H., and ONUFRYK, P.Z.: Network

Processor Design : Issues and Practices, Vol. 1, San Francisco: Morgan Kaufmann, 2002.

DALLY, W.J., and TOWLES, B.P.: Principles and Practices of Interconnection Networks,

San Francisco: Morgan Kaufmann, 2004.

DANESHBEH, A.K., and HASAN, M.A.: ‘‘Area Efficient High Speed Elliptic Curve Cryp-

toprocessor for Random Curves,’’ Proc. Int’l Conf. on Inf. Tech.: Codingspeculative xecution and Computing, IEEE, pp. 588-593, 2004.

DEAN, A.G.: ‘‘Efficient Real-Time Fine-Grained Concurrency on Low-Cost Microcon-

trollers,’’ IEEE Micro Magazine, vol. 24, pp. 10-22, July-Aug. 2004.

DENNING, P.J.: ‘‘The Working Set Model for Program Behavior,’’ Commun. of the ACM,

vol. 11, pp. 323-333, May 1968.

DIJKSTRA, E.W.: ‘‘GOTO Statement Considered Harmful,’’ Commun. of the ACM, vol.

11, pp. 147-148, March 1968a.

DIJKSTRA, E.W.: ‘‘Co-operating Sequential Processes,’’ in Programming Languages, F.

Genuys (ed.), New York: Academic Press, 1968b.

DONALDSON, G., and JONES, D.: ‘‘Cable Television Broadband Network Architectures,’’

IEEE Commun. Magazine, vol. 39, pp. 122-126, June 2001.

668

READING LIST AND BIBLIOGRAPHY

CHAP. 9

DONGARRA, J., FOSTER, I., FOX, G., GROPP, W., KENNEDY, K., TORCZON, L., and WHITE, A.: The Sourcebook of Parallel Computing. San Francisco: Morgan Kauf-

man, 2003.

DUBOIS, M., SCHEURICH, C., and BRIGGS, F.A.: ‘‘Memory Access Buffering in Mul-

tiprocessors,’’ Proc. 13th Ann. Int’l Symp. on Computer Arch., ACM, pp. 434-442, 1986.

DULONG, C.: ‘‘The IA-64 Architecture at Work,’’ IEEE Computer Magazine, vol. 31, pp.

24-32, July 1998.

DUTTA-ROY, A.: ‘‘An Overview of Cable Modem Technology and Market Perspectives,’’

IEEE Commun. Magazine, vol. 39, pp. 81-88, June 2001.

ESCHMANN, F., KLAUER, B., MOORE, R., and WALDSCHMIDT, K.: ‘‘SDAARC: An

Extended Cache-Only Memory Architecture,’’ IEEE Micro Magazine, vol. 22, pp. 62-70, May-June, 2002.

FAGGIN, F., HOFF, M.E., Jr., MAZOR, S., and SHIMA, M.: ‘‘The History of the 4004,’’

IEEE Micro Magazine, vol. 16, pp. 10-20, Nov. 1996.

FALCON, A., STARK, J., RAMIREZ, A., LAI, K., and VALERO, M.: ‘‘Prophet/Critic Hybrid Branch Prediction,’’ Proc. 31th Ann. Int’l Symp. on Computer Arch., ACM, pp. 250-261, 2004. FISHER, J.A., and FREUDENBERGER, S.M.: ‘‘Predicting Conditional Branch Directions

from Previous Runs of a Program,’’ Proc. Fifth Int’l Conf. on Arch. Support for Prog. Lang. and Operating Syst., ACM, pp. 85-95, 1992.

FLOYD, T.L.: Digital Fundamentals, 8th ed., Upper Saddle River, NJ: Prentice Hall, 2002. FLYNN, D.: ‘‘AMBA: Enabling Reusable On-Chip Designs,’’ IEEE Micro Magazine, vol.

17, pp. 20-27, July 1997.

FLYNN, M.J.: ‘‘Some Computer Organizations and Their Effectiveness,’’ IEEE Trans. on

Computers, vol. C-21, pp. 948-960, Sept. 1972.

FOSTER, I., and KESSELMAN, C.: The Grid 2: Blueprint for a New Computing Infrastruc-

ture, San Francisco: Morgan Kaufman, 2003.

FOSTER, I., KESSELMAN, C., NICK, J.M., and TUECKE, S.: ‘‘Grid Services for Distri-

buted Systems Integration,’’ IEEE Computer Magazine, vol. 35, pp. 37-46, June 2002.

FOSTER, I., and KESSELMAN, C.: ‘‘Globus: A Metacomputing Infrastructure Toolkit,’’

Int’l J. of Supercomputer Applications, vol. 11, pp. 115-128, 1998a.

FOSTER, I., and KESSELMAN, C.: ‘‘The Globus Project: A Status Report,’’ IPPS/SPDP

’98 Heterogeneous Computing Workshop, IEEE, pp. 4-18, 1998b.

FOTHERINGHAM, J.: ‘‘Dynamic Storage Allocation in the Atlas Computer Including an

Automatic Use of a Backing Store,’’ Commun. of the ACM, vol. 4, pp. 435-436, Oct. 1961.

SEC. 9.2

ALPHABETICAL BIBLIOGRAPHY

669

GEIST, A., BEGUELIN, A., DONGARRA, J., JIANG, W., MANCHECK, R., and SUNDERRAM, V.: PVM: Parallel Virtual Machine — A User’s Guide and Tutorial for

Networked Parallel Computing, Cambridge, MA: M.I.T. Press, 1994.

GERBER, R., and BINSTOCK, A.: Programming with Hyper-Threading Technology, Santa

Clara, CA: Intel Press, 2004.

GHEMAWAT, S., GOBIOFF, H., and LEUNG, S.-T.: ‘‘The Google File System,’’ Proc.

19th Symp. on Operating Systems Principles, ACM, pp. 29-43, 2003.

GOODMAN, J.R.: ‘‘Using Cache Memory to Reduce Processor Memory Traffic,’’ Proc.

10th Ann. Int’l Symp. on Computer Arch., ACM, pp. 124-131, 1983.

GOODMAN, J.R.: ‘‘Cache Consistency and Sequential Consistency,’’ Tech. Rep. 61, IEEE

Scalable Coherent Interface Working Group, IEEE, 1989.

GRIMSHAW, A.S., and WULF, W.: ‘‘Legion: A View from 50,000 Feet,’’ Proc. Fifth Int’l

Symp. on High-Performance Distributed Computing, IEEE, pp. 89-99, Aug. 1996.

GRIMSHAW, A.S., and WULF, W.: ‘‘The Legion Vision of a Worldwide Virtual Com-

puter,’’ Commun. of the ACM, vol. 40, pp. 39-45, Jan. 1997.

GROPP, W., LUSK, E, and SKJELLUM, A: ‘‘Using MPI: Portable Parallel Programming

with the Message Passing Interface,’’ Cambridge, MA: M.I.T. Press, 1994.

GURUMURTHI, S., SIVASUBRAMANIAM, KANDEMIR, M., and FRANKE, H.: ‘‘Reduc-

ing Disk Power Consumption in Servers with DRPM,’’ IEEE Computer Magazine, vol. 36, pp. 59-66, Dec. 2003.

HAGERSTEN, E., LANDIN, A., HARIDI, S.: ‘‘DDM—A Cache-Only Memory Architec-

ture,’’ IEEE Computer Magazine, vol. 25, pp. 44-54, Sept. 1992.

HAMACHER, V.V., VRANESIC, Z.G., and ZAKY, S.G.: Computer Organization, 5th ed.,

New York: McGraw-Hill, 2001.

HAMMING, R.W.: ‘‘Error Detecting and Error Correcting Codes,’’ Bell Syst. Tech. J., vol.

29, pp. 147-160, April 1950.

HAMMOND, L., WONG, V., CHEN, M., HERTZBERG, B, DAVIS, J., CARLSTROM, B., PRABHU, M., WIJAYA, H., KOZYRAKIS, C., and OLUKOTUN, K.: ‘‘Transactional

Memory Coherence and Consistency,’’ Proc. 31th Ann. Int’l Symp. on Computer Arch., ACM, pp. 102-113, 2004.

HANDY, J.: The Cache Memory Book, 2nd ed., Orlando, FL: Academic Press, 1998. HART, J.M.: Win32 System Programming, Reading, MA: Addison-Wesley, 1997. HEATH, S.: Embedded Systems Design, Oxford: Newnes, 2003. HENKEL, J., HU, X.S., and BHATTACHARYYA, S.S.: ‘‘Taking on the Embedded System

Challenge,’’ IEEE Computer Magazine, vol. 36, pp. 35-37, April 2003.

HENNESSY, J.L.: ‘‘VLSI Processor Architecture,’’ IEEE Trans. on Computers, vol. C-33,

pp. 1221-1246, Dec. 1984.

670

READING LIST AND BIBLIOGRAPHY

CHAP. 9

HENNESSY, J.L., and PATTERSON, D.A.: Computer Architecture A Quantitative

Approach, 3rd ed. San Francisco: Morgan Kaufmann, 2003.

HILL, M.: ‘‘Multiprocessors Should Support Simple Memory-Consistency Models,’’ IEEE

Computer Magazine, vol. 31, pp. 28-34, Aug. 1998.

HINTON, G., SAGER, D., UPTON, M., BOGGS, D., CARMEAN, D., KYKER, A., ROUSSEL, P.: ‘‘The Microarchtecture of the Pentium 4,’’ Intel Technology Journal,

vol. 5, pp. 1-12, Jan.-March, 2001.

HOARE, C.A.R.: ‘‘Monitors, An Operating System Structuring Concept,’’ Commun. of the

ACM, vol. 17, pp. 549-557, Oct. 1974; Erratum in Commun. of the ACM, vol. 18, p. 95, Feb. 1975.

HUH, J., BURGER, D., CHANG, J., and SOHI, G.S.: ‘‘ Speculative Incoherent Cache Proto-

cols,’’ IEEE Micro Magazine, vol. 24, pp. 104-109, Nov.-Dec. 2004.

HWANG, K., and XU, Z.: Scalable Parallel Computing, New York: McGraw-Hill, 1998. HWU, W.-M.: ‘‘Introduction to Predicated Execution,’’ IEEE Computer Magazine, vol. 31,

pp. 49-50, Jan. 1998.

IRVINE, K: Assembly Language for Intel-Based Computers, 4th ed., Upper Saddle River,

NJ: Prentice Hall, 2002.

JACOB, B., and MUDGE, T.: ‘‘Virtual Memory: Issues of Implementation,’’ IEEE Com-

puter Magazine, vol. 31, pp. 33-43, June 1998a.

JACOB, B., and MUDGE, T.: ‘‘Virtual Memory in Contemporary Microprocessors,’’ IEEE

Micro Magazine, vol. 18, pp. 60-75, July-Aug. 1998b.

JERRAYA, A.A., and WOLF. W.: Multiprocessor Systems-on-a-Chip, San Francisco: Mor-

gan Kaufmann, 2005.

JIMENEZ, D.A.: ‘‘Fast Path-Based Neural Branch Prediction,’’ Proc. 36th Int’l Symp. on

Microarchitecture, IEEE., pp. 243-252, 2003.

JOHNSON, K.L., KAASHOEK, M.F., and WALLACH, D.A.: ‘‘CRL: High-Performance

All-Software Distributed Shared Memory,’’ Proc. 15th Symp. on Operating Systems Principles, ACM, pp. 213-228, 1995.

JOHNSON, M.: Superscalar Microprocessor Design, Upper Saddle River, NJ: Prentice

Hall, 1991..

KALLA, R., SINHAROY, B., and TENDLER, J.M.: ‘‘IBM Power5 Chip: A Dual-Core Mul-

tithreaded Processor,’’ IEEE Micro Magazine, vol. 24, pp. 40-47, March-April 2004.

KAPASI, U.J., RIXNER, S., DALLY, W.J., KHAILANY, B., AHN, J.H., MATTSON, P., and OWENS, J.D.: ‘‘Programmable Stream Processors,’’ IEEE Computer Magazine, vol.

36, pp. 54-62, Aug. 2003.

KAPIL, S., MCGHAN, H., and LAWRENDRA, J.: ‘‘A Chip Multithreaded Processor for

Network-Facing Workloads,’’ IEEE Micro Magazine, vol. 24, pp. 20-30, March-April 2004.

SEC. 9.2

ALPHABETICAL BIBLIOGRAPHY

671

KATZ, R.H., and BORRIELLO, G.: Contemporary Logic Design, Upper Saddle River, NJ:

Prentice Hall, 2004.

KAUFMAN, C., PERLMAN, R., and SPECINER, M.: Network Security, 2nd ed., Upper

Saddle River, NJ: Prentice Hall, 2002.

KERMARREC., A.-M., KUZ, I., VAN STEEN, M., and TANENBAUM, A.S.: ‘‘A Framework

for Consistent Replicated Web Objects,’’ Proc. 18th Int’l Conf. on Distr. Computing Syst., IEEE, pp. 276-284, 1998.

KIM, N.S., AUSTIN, T., BLAAUW, D., MUDGE, T., FLAUTNER, K., HU, J.S., IRWIN, M.J., KANDEMIR, M., and NARAYANAN, V.: ‘‘Leakage Current: Moore’s Law Meets

Static Power,’’ IEEE Computer Magazine, vol. 36, 68-75, Dec. 2003.

KNUTH, D.E.: ‘‘An Empirical Study of FORTRAN Programs,’’ Software—Practice &

Experience, vol. 1, pp. 105-133, 1971.

KNUTH, D.E.: The Art of Computer Programming: Fundamental Algorithms, 3rd ed.,

Reading, MA: Addison-Wesley, 1997.

KNUTH, D.E.: The Art of Computer Programming: Seminumerical Algorithms, 3rd ed.,

Reading, MA: Addison-Wesley, 1998.

KOGEL, T., and MYER, H.: ‘‘Heterogeneous MP-SoC: the solution to energy-efficient sig-

nal processing,’’ Proc. 41st Ann. Conf. on Design Automation, IEEE, pp. 686-691, 2004.

KONTOTHANASSIS, L., HUNT, G., STETS, R., HARDAVELLAS, N., CIERNIAD, M., PARTHASARATHY, S., MEIRA, W., DWARKADAS, S., and SCOTT, M.: VM-Based

Shared Memory on Low Latency Remote Memory Access Networks, Proc. 24th Ann. Int’l Symp. on Computer Arch., ACM, pp. 157-169, 1997.

KOREN, I.: Computer Arithmetic Algorithms, Natick, MA:A.K. Peters, 2002. KOUFATY, D., and MARR, D. T.: ‘‘Hyperthreading Technology in the Netburst Microar-

chitecture,’’ IEEE Micro Magazine, vol. 23, pp. 56-65, March-April 2003.

KUMAR, R., JOUPPI, N.P., and TULLSEN, D.M.: ‘‘Conjoined-Core Chip Multiprocess-

ing,’’ Proc. 37th Int’l Symp. on Microarchitecture, IEEE., pp. 195-206, 2004.

LAMPORT, L.: ‘‘How to Make a Multiprocessor Computer That Correctly Executes Mul-

tiprocess Programs,’’ IEEE Trans. on Computers, vol. C-28, pp. 690-691, Sept. 1979.

LAROWE, R.P., and ELLIS, C.S.: ‘‘Experimental Comparison of Memory Management

Policies for NUMA Multiprocessors,’’ ACM Trans. on Computer Systems, vol. 9, pp. 319-363, Nov. 1991.

LAVAGNO, L.: ‘‘Systems on a Chip: The Next Electronic Frontier,’’ IEEE Micro Maga-

zine, vol. 22, pp. 14-15, Sept.-Oct. 2002.

LAWTON, G.: ‘‘Will Network Processor Units Live up to Their Promise?,’’ IEEE Com-

puter Magazine, vol. 37, pp. 13-15, April 2004.

672

READING LIST AND BIBLIOGRAPHY

CHAP. 9

LEKKAS, P.C.: Network Processors: Architectures, Protocols, and Platforms, New York:

McGraw-Hill, 2003.

LEVINE, J.R.: Linkers and Loaders, San Francisco: Morgan Kaufmann, 2000. LI, K., and HUDAK, P.: ‘‘Memory Coherence in Shared Virtual Memory Systems,’’ ACM

Trans. on Computer Systems, vol. 7, pp. 321-359, Nov. 1989.

LIMA, F., CARRO, L., VELAZCO, R., and REIS, R.: ‘‘Injecting Multiple Upsets in a SEU

Tolerant 8051 Micro-Controller,’’ Proc. Eighth IEEE Int’l On-Line Testing Workshop IEEE, p. 194, July 2002.

LINES, A.: ‘‘Asynchronous Interconnect for Synchronous SoC Design,’’ IEEE Micro

Magazine, vol. 24, pp. 32-41, Jan.-Feb. 2004.

LU, H., COX, A.L., DWARKADAS, S., RAJAMONY, R., and ZWAENEPOEL, W.:

‘‘Software Distributed Shared Memory Support for Irregular Applications,’’ Proc. Sixth Conf. on Prin. and Practice of Parallel Progr., pp. 48-56, June 1997.

LUKASIEWICZ, J.: Aristotle’s Syllogistic, 2nd ed., Oxford: Oxford University Press,

1958.

LUTZ, J., and HASAN, A.: ‘‘High Performance FPGA based Elliptic Curve Cryptographic

Co-Processor,’’ Proc. Int’l Conf. on Inf. Tech.: Coding and Computing, IEEE, pp. 486-492, 2004.

LYYTINEN, K., and Yoo, Y.: ‘‘Issues and Challenges in Ubiquitous Computing,’’ Commun.

of the ACM, vol. 45, pp. 63-65, Dec. 2002.

MACKENZIE, I.S., PHAN, R.: The 8051 Microcontroller, 4th ed. Upper Saddle River, NJ:

Prentice Hall, 2005.

MANO, M.M., and KIME, C.R.: Logic and Computer Design Fundamentals, 3rd ed.,

Upper Saddle River, NJ: Prentice Hall, 2003.

MARTIN, A.J., NYSTROM, M., PAPADANTONAKIS, K., PENZES, P.I., PRAKASH, P., WONG, C.G., CHANG, J., KO, K.S., LEE, B., OU, E., PUGH, J, TALVALA, E-V., TONG, J.T., and TURA, A.: ‘‘The Lutonium: A Sub-Nanojoule Asynchronous 8051

Microcontroller,’’ Proc. Ninth Int’l Symp. on Asynchronous Circuits and Systems IEEE, pp. 14-23, 2003.

MARTIN, R.P., VAHDAT, A.M., CULLER, D.E., and ANDERSON, T.E.: ‘‘Effects of Com-

munication Latency, Overhead, and Bandwidth in a Cluster Architecture,’’ Proc. 24th Ann. Int’l Symp. on Computer Arch., ACM, pp. 85-97, 1997.

MAYHEW, D., and KRISHNAN, V.: ‘‘ PCI Express and Advanced Switching: Evolutionary

Path to Building Next Generation Interconnects,’’ Proc. 11th Symp. on High Perf. Interconnects IEEE, pp. 21-29, Aug. 2003.

MAZIDI, M.A., MCKINLAY, and MAZIDI, J.G.: 8051 Microcontroller and Embedded Sys-

tems Upper Saddle River, NJ: Prentice Hall, 2005.

SEC. 9.2

ALPHABETICAL BIBLIOGRAPHY

673

MAZIDI, M.A., and MAZIDI, J.G.: The 80x86 IBM PC and Compatible Computers, 4th

ed., Upper Saddle River, NJ: Prentice Hall, 2002.

MCKNIGHT, L.W., HOWISON, J., and BRADNER, S: ‘‘Wireless Grids,’’ IEEE Internet

Computing, vol. 8, pp. 24-31, July-Aug. 2004.

MCKUSICK, M.K., BOSTIC, K., KARELS, M., and QUARTERMAN, J.S.: ‘‘The Design and

Implementation of the 4.4 BSD Operating System,’’ Reading, MA: Addison-Wesley, 1996.

MCKUSICK, M.K., JOY, W.N., LEFFLER, S.J., and FABRY, R.S.: ‘‘A Fast File System for

UNIX,’’ ACM Trans. on Computer Systems, vol. 2, pp. 181-197, Aug. 1984.

MCNAIRY, C., and SOLTIS, D.: ‘‘Itanium 2 Processor Microarchitecture,’’ IEEE Micro

Magazine, vol. 23, pp. 44-55, March-April 2003.

MIN, R., Jone, W.-Ben., and HU, Y.: ‘‘Location Cache: A Low-Power L2 Cache System,’’

Proce. 2004 Int’l Symp. on Low Power Electronics and Design, IEEE, pp. 120-125, Aug. 2004.

MESSMER, H.-P.: The Indispensible PC Hardware Book, 4th ed., Reading, MA:

Addison-Wesley, 2001.

MOUDGILL, M., and VASSILIADIS, S.: ‘‘Precise Interrupts,’’ IEEE Micro Magazine, vol.

16, pp. 58-67, Jan. 1996.

MULLENDER, S.J., and TANENBAUM, A.S.: ‘‘Immediate Files,’’ Software—Practice and

Experience, vol. 14, pp. 365-368, 1984.

NESBIT, K.J., and SMITH, J.E.: ‘‘Data Cache Prefetching Using a Global History Buffer,’’

Proc. 10th Int’l Symp. on High Perf. Computer Arch.,’’ IEEE, pp. 96-106, 2004.

NG, S.W.: ‘‘Advances in Disk Technology: Performance Issues,’’ IEEE Computer Maga-

zine, vol. 31, pp. 75-81, May 1998.

NICKOLLS, J., MADAR, L.J. III, Johnson, S., RUSTAGI, V., UNGER, K., and CHOUDHURY, M.: ‘‘Calisto: A Low-Power Single-Chip Multiprocessor Communica-

tions Platform,’’ IEEE Micro Magazine, vol. 23, pp. 29-43, March 2003.

NORTON, P., and GOODMAN, J.: Inside the PC, 8th ed., Indianapolis, IN: Sams, 1999. NULL, L., and LOBUR, J.: The Essentials of Computer Organization and Architecture,

Sudbury, MA: Jones and Bartlett, 2003.

O’CONNOR, J.M., and TREMBLAY, M.: ‘‘PicoJava-I: The Java Virtual Machine in

Hardware,’’ IEEE Micro Magazine, vol. 17, pp. 45-53, March-April 1997.

ORGANICK, E.: The MULTICS System, Cambridge, MA: M.I.T. Press, 1972. OSKIN, M., CHONG, F.T., and CHUANG, I.L.: ‘‘A Practical Architecture for Reliable

Quantum Computers,’’ IEEE Computer Magazine, vol. 35, pp. 79-87, Jan. 2002.

OUADJAOUT, S., and HOUZET, D.: ‘‘Easy SoC Design with VCI SystemC Adapters,’’

Proc. Digital System Design, IEEE, pp. 316-323, 2004.

674

READING LIST AND BIBLIOGRAPHY

CHAP. 9

PAPAEFSTATHIOU, I., NIKOLAOU, N,A, DOSHI, B., and GROSSE, E.: ‘‘Network Proces-

sors for Future High-End Systems and Applications,’’ IEEE Micro Magazine, vol. 24, pp. 7-9, Sept.-Oct. 2004.

PAPAMARCOS, M., and PATEL., J.: ‘‘A Low Overhead Coherence Solution for Multipro-

cessors with Private Cache Memories,’’ Proc. 11th Ann. Int’l Symp. on Computer Arch., ACM, pp. 348-354, 1984.

PARIKH, D., SKADRON, K., ZHANG, Y., and STAN, M.: ‘‘Power-Aware Branch Predic-

tion: Characterization and Design,’’ IEEE Trans. on Computers, vol. 53, 168-186, Feb. 2004.

PATTERSON, D.A.: ‘‘Reduced Instruction Set Computers,’’ Commun. of the ACM, vol. 28,

pp. 8-21, Jan. 1985.

PATTERSON, D.A., GIBSON, G., and KATZ, R.: ‘‘A case for redundant arrays of inexpen-

sive disks (RAID),’’ Proc. ACM SIGMOD Int’l Conf. on Management of Data, ACM, pp. 109-166,1988.

PATTERSON, D.A., and HENNESSY, J.L.: Computer Organization and Design, 3rd ed.,

San Francisco: Morgan Kaufmann, 2005.

PATTERSON, D.A., and SEQUIN, C.H.: ‘‘A VLSI RISC,’’ IEEE Computer Magazine, vol.

15, pp. 8-22, Sept. 1982.

PAUL, R.P.: SPARC Architecture, Assembly Language, Programming, and C, Upper Sad-

dle River, NJ: Prentice Hall, 1994.

PFISTER, G.F.: In Search of Clusters, 2nd ed., Upper Saddle River, NJ: Prentice Hall,

1998.

POPESCU, B.C., STEEN, M. VAN, and TANENBAUM, A.S.: ‘‘A Security Architecture for

Object-Based Distributed Systems,’’ Proc. 18th Annual Computer Security Appl. Conf., ACM, pp. 161-171, 2002.

POUNTAIN, D.: ‘‘Pentium: More RISC than CISC,’’ Byte, vol. 18, pp. 195-204, Sept.

1993.

PRICE, D.: ‘‘A History of Calculating Machines,’’ IEEE Micro Magazine, vol. 4, pp. 22-

52, Jan. 1984.

RADIN, G.: ‘‘The 801 Minicomputer,’’ Computer Arch. News, vol. 10, pp. 39-47, March

1982.

RAMAN, S.K., PENTKOVSKI, V., and KESHAVA, J.: ‘‘Implementing Streaming SIMD

Extensions on the Pentium III Processor,’’ IEEE Micro Magazine, vol. 20, pp. 47-57, July-Aug. 2000.

RAVIKUMAR, C.P.: ‘‘Multiprocessor Architectures for Embedded System-on-a-Chip

Applications,’’ Proc. 17th Int’l Conf. on VLSI Design, IEEE, pp. 512-519, Jan. 2004.

RITCHIE, D.M., and THOMPSON, K.: ‘‘The UNIX Time-Sharing System,’’ Commun. of

the ACM, vol. 17, pp. 365-375, July 1974.

SEC. 9.2

ALPHABETICAL BIBLIOGRAPHY

675

ROBINSON, G.S.: ‘‘Toward the Age of Smarter Storage,’’ IEEE Computer Magazine, vol.

35, pp. 35-41, Dec. 2002.

ROSENBLUM, M., and OUSTERHOUT, J.K.: ‘‘The Design and Implementation of a Log-

Structured File System,’’ Proc. Thirteenth Symp. on Operating System Principles, ACM, pp. 1-15, 1991.

ROTH, C.H.: Fundamentals of Logic Design, 5th ed., Florence, KY:Thomson Engineering,

2003.

RUSSINOVICH, M.E., and SOLOMON, D.A.: Microsoft Windows Internals, 4th ed., Red-

mond, WA: Microsoft Press, 2005.

RUSU, S., MULJONO, H., and CHERKAUER, B.: ‘‘Itanium 2 Processor 6M,’’ IEEE Micro

Magazine, vol. 24, pp. 10-18, March-April 2004.

SAHA, D., and MUKHERJEE, A.: ‘‘Pervasive Computing: A Paradigm for the 21st Cen-

tury,’’ IEEE Computer Magazine, vol. 36, pp. 25-31, March 2003.

SAKAMURA, K: ‘‘Making Computers Invisible,’’ IEEE Micro Magazine, vol. 22, pp. 7-

11, 2002.

SALOMAN, D.: Assemblers and Loaders, Upper Saddle River, NJ: Prentice Hall, 1993. SCALES, D.J., GHARACHORLOO, K., and THEKKATH, C.A.: ‘‘Shasta: A Low-Overhead

Software-Only Approach for Supporting Fine-Grain Shared Memory,’’ Proc. Seventh Int’l Conf. on Arch. Support for Prog. Lang. and Oper. Syst., ACM, pp. 174-185, 1996.

SCHEIBLE, J.P.: ‘‘A Survey of Storage Options,’’ IEEE Computer Magazine, vol. 35, pp.

42-46, Dec. 2002.

SELTZER, M., BOSTIC, K., MCKUSICK, M.K., and STAELIN, C.: ‘‘An Implementation of

a Log-Structured File System for UNIX,’’ Proc. Winter 1993 USENIX Technical Conf., pp. 307-326, 1993.

SHANLEY, T., and ANDERSON, D.: PCI System Architecture, 4th ed., Reading, MA:

Addison-Wesley, 1999.

SHRIVER, B., and SMITH, B.: The Anatomy of a High-Performance Microprocessor: A

Systems Perspective, Los Alamitos, CA: IEEE Computer Society, 1998.

SIMA, D: ‘‘Superscalar Instruction Issue,’’ IEEE Micro Magazine, vol. 17, pp. 28-39,

Sept.-Oct 1997.

SIMA, D., FOUNTAIN, T., and KACSUK, P.: Advanced Computer Architectures: A Design

Space Approach, Reading, MA: Addison-Wesley, 1997.

SLATER, R.: Portraits in Silicon, Cambridge, MA: M.I.T. Press, 1987. SOHI, G.S., and ROTH, A.: ‘‘Speculative Multithreaded Processors,’’ IEEE Computer

Magazine, vol. 34, pp. 66-73, April 2001.

676

READING LIST AND BIBLIOGRAPHY

CHAP. 9

SNIR, M., OTTO, S.W., HUSS-LEDERMAN, S., WALKER, D.W., and DONGARRA, J.:

MPI: The Complete Reference Manual, Cambridge, MA: M.I.T. Press, 1996.

SOLARI, E., and CONGDON, B.: PCI Express Design & System Architecture, Research

Tech, INc., 2005.

SOLARI, E., and WILLSE, G.: PCI and PCI-X Hardware and Software, 6th ed., San

Diego, CA: Annabooks, 2004.

STALLINGS, W.: Computer Organization and Architecture, 6th ed., Upper Saddle River,

NJ: Prentice Hall, 2003.

STENSTROM, P., HAGERSTEN, E., LILJA, D.J., MARTONOSI, M., and VENUGOPAL, M.: ‘‘Trends in Shared Memory Multiprocessing,’’ IEEE Computer Magazine, vol. 30,

pp. 44-50, Dec. 1997.

STETS, R., DWARKADAS, S., HARDAVELLAS, N., HUNT, G., KONTOTHANASSIS, L., PARTHASARATHY, S., and SCOTT, M.: ‘‘CASHMERE-2L: Software Coherent

Shared Memory on Clustered Remote-Write Networks,’’ Proc. 16th Symp. on Operating Systems Principles, ACM, pp. 170-183, 1997.

SUH, T., LEE, H.-H. S., BLOUGH, D.M.: ‘‘Integrating Cache Coherence Protocols for

Heterogeneous Multiprocessor Systems, Part 1,’’ IEEE Micro Magazine, vol. 24, pp. 33-41, July 2004.

SUMMERS, C.K.: ADSL: Standards, Implementation, and Architecture, Boca Raton, FL:

CRC Press, 1999.

SUNDERRAM, V.B..: ‘‘PVM: A Framework for Parallel Distributed Computing,’’ Con-

currency: Practice and Experience, vol. 2, pp. 315-339, Dec. 1990.

SWAN, R.J., FULLER, S.H., and SIEWIOREK, D.P.: ‘‘Cm*—A Modular Multiprocessor,’’

Proc. NCC, pp. 645-655, 1977.

TAN, W.M.: Developing USB PC Peripherals, San Diego, CA: Annabooks, 1997. TANENBAUM, A.S.: ‘‘Computer Networks,’’ Upper Saddle River, NJ: Prentice Hall, 2003. TANENBAUM, A.S.: ‘‘Implications of Structured Programming for Machine Architec-

ture,’’ Commun. of the ACM, vol. 21, pp. 237-246, March 1978.

TANENBAUM, A.S.: Operating Systems: Design and Implementation, Upper Saddle River,

NJ: Prentice Hall, 1987.

TANENBAUM, A.S, and WOODHULL, A.W.: Operating Systems: Design and Implementa-

tion, 2nd ed., Upper Saddle River, NJ: Prentice Hall, 1997.

THOMPSON, K.: ‘‘UNIX Implementation,’’ Bell Syst. Tech. J., vol. 57, pp. 1931-1946,

July-Aug. 1978.

TRELEAVEN, P.: ‘‘Control-Driven, Data-Driven, and Demand-Driven Computer Archi-

tecture,’’ Parallel Computing, vol. 2, 1985.

SEC. 9.2

ALPHABETICAL BIBLIOGRAPHY

677

TREMBLAY, M., and O’CONNOR, J.M.: ‘‘UltraSPARC I: A Four-Issue Processor Sup-

porting Multimedia,’’ IEEE Micro Magazine, vol. 16, pp. 42-50, March 1996.

TRIEBEL, W.A.: The 80386, 80486, and Pentium Processor, Upper Saddle River, NJ:

Prentice Hall, 1998.

TUCK, N., and TULLSEN, D.M.: ‘‘Initial Observations of the Simultaneous Multithreading

Pentium 4 Processor,’’ Proc. 12th Int’l Conf. on Parallel Arch. and Compilation Techniques, IEEE, pp. 26-35, 2003.

UNGER, S.H.: ‘‘A Computer Oriented Toward Spatial Problems,’’ Proc. IRE, vol. 46, pp.

1744-1750, 1958.

VAHALIA, U.: UNIX Internals, Upper Saddle River, NJ: Prentice Hall, 1996. VAHID, F.: ‘‘The Softening of Hardware,’’ IEEE Computer Magazine, vol. 36, pp. 27-34,

April 2003.

VAN STEEN, M., HOMBURG, P.C., and TANENBAUM, A.S.: ‘‘The Architectural Design

of Globe: A Wide-Area Distributed System,’’ IEEE Concurrency, vol. 7, pp. 70-78, Jan.-March 1999.

VETTER, P., GODERIS, D., VERPOOTEN, L., and GRANGER, A.: ‘‘Systems Aspects of

APON/VDSL Deployment,’’ IEEE Commun. Magazine, vol. 38, pp. 66-72, May 2000.

WEAVER, D.L., and GERMOND, T.: The SPARC Architecture Manual, Version 9, Upper

Saddle River, NJ: Prentice Hall, 1994.

WEISER, M.: ‘‘The Computer for the 21st Century,’’ IEEE Pervasive Computing, vol. 1,

pp. 19-25, Jan.-March 2002; originally published in Scientific American, Sept. 1991.

WILKES, M.V.: ‘‘Computers Then and Now,’’ J. ACM, vol. 15, pp. 1-7, Jan. 1968. WILKES, M.V.: ‘‘The Best Way to Design an Automatic Calculating Machine,’’ Proc.

Manchester Univ. Computer Inaugural Conf., 1951.

WILSON, J.: ‘‘Challenges and Trends in Processor Design,’’ IEEE Computer Magazine,

vol. 31, pp. 39-48, Jan. 1998.

WILSON, P.: ‘‘Floating-Point Survival Kit,’’ Byte, vol. 13, pp. 217-226, March 1988. WOLF, W.: ‘‘The Future of Multiprocessor Systems-on-Chips,’’ Proc. 41st Ann. Conf. on

Design Automation, IEEE, pp. 681-685, 2004.

This page intentionally left blank

A BINARY NUMBERS

The arithmetic used by computers differs in some ways from the arithmetic used by people. The most important difference is that computers perform operations on numbers whose precision is finite and fixed. Another difference is that most computers use the binary rather than the decimal system for representing numbers. These topics are the subject of this appendix.

A.1 FINITE-PRECISION NUMBERS While doing arithmetic, one usually gives little thought to the question of how many decimal digits it takes to represent a number. Physicists can calculate that there are 1078 electrons in the universe without being bothered by the fact that it requires 79 decimal digits to write that number out in full. Someone calculating the value of a function with pencil and paper who needs the answer to six significant digits simply keeps intermediate results to seven, or eight, or however many are needed. The problem of the paper not being wide enough for seven-digit numbers never arises. With computers, matters are quite different. On most computers, the amount of memory available for storing a number is fixed at the time that the computer is designed. With a certain amount of effort, the programmer can represent numbers two, or three, or even many times larger than this fixed amount, but doing so does not change the nature of this difficulty. The finite nature of the computer forces 679

680

BINARY NUMBERS

APP. A

us to deal only with numbers that can be represented in a fixed number of digits. We call such numbers finite-precision numbers. In order to study properties of finite-precision numbers, let us examine the set of positive integers representable by three decimal digits, with no decimal point and no sign. This set has exactly 1000 members: 000, 001, 002, 003, ..., 999. With this restriction, it is impossible to express certain kinds of numbers, such as 1. Numbers larger than 999. 2. Negative numbers. 3. Fractions. 4. Irrational numbers. 5. Complex numbers. One important property of arithmetic on the set of all integers is closure with respect to the operations of addition, subtraction, and multiplication. In other words, for every pair of integers i and j, i + j, i − j, and i × j are also integers. The set of integers is not closed with respect to division, because there exist values of i and j for which i / j is not expressible as an integer (e.g., 7/2 and 1/0). Finite-precision numbers are not closed with respect to any of these four basic operations, as shown below, using three-digit decimal numbers as an example: 600 + 600 = 1200 003 − 005 = −2 050 × 050 = 2500 007 / 002 = 3.5

(too large) (negative) (too large) (not an integer)

The violations can be divided into two mutually exclusive classes: operations whose result is larger than the largest number in the set (overflow error) or smaller than the smallest number in the set (underflow error), and operations whose result is neither too large nor too small but is simply not a member of the set. Of the four violations above, the first three are examples of the former, and the fourth is an example of the latter. Because computers have finite memories and therefore must of necessity perform arithmetic on finite-precision numbers, the results of certain calculations will be, from the point of view of classical mathematics, just plain wrong. A calculating device that gives the wrong answer even though it is in perfect working condition may appear strange at first, but the error is a logical consequence of its finite nature. Some computers have special hardware that detects overflow errors. The algebra of finite-precision numbers is different from normal algebra. As an example, consider the associative law: a + (b − c) = (a + b ) − c Let us evaluate both sides for a = 700, b = 400, c = 300. To compute the lefthand side, first calculate (b − c), which is 100, and then add this amount to a,

SEC. A.1

681

FINITE-PRECISION NUMBERS

yielding 800. To compute the right-hand side, first calculate (a + b), which gives an overflow in the finite arithmetic of three-digit integers. The result may depend on the machine being used but it will not be 1100. Subtracting 300 from some number other than 1100 will not yield 800. The associative law does not hold. The order of operations is important. As another example, consider the distributive law: a × (b − c) = a × b − a × c Let us evaluate both sides for a = 5, b = 210, c = 195. The left-hand side is 5 × 15, which yields 75. The right-hand side is not 75 because a × b overflows. Judging from these examples, one might conclude that although computers are general-purpose devices, their finite nature renders them especially unsuitable for doing arithmetic. This conclusion is, of course, not true, but it does serve to illustrate the importance of understanding how computers work and what limitations they have.

A.2 RADIX NUMBER SYSTEMS An ordinary decimal number with which everyone is familiar consists of a string of decimal digits and, possibly, a decimal point. The general form and its usual interpretation are shown in Fig. A-1. The choice of 10 as the base for exponentiation, called the radix, is made because we are using decimal, or base 10, numbers. When dealing with computers, it is frequently convenient to use radices other than 10. The most important radices are 2, 8, and 16. The number systems based on these radices are called binary, octal, and hexadecimal, respectively.

dn

…

100's place

10's place

1's place

d2

d1

d0

. n

Number =

Σ

i = –k

.1's place

.01's place

.001's place

d–1

d–2

d–3

…

d–k

di × 10i

Figure A-1. The general form of a decimal number.

A radix k number system requires k different symbols to represent the digits 0 to k − 1. Decimal numbers are built up from the 10 decimal digits

682

BINARY NUMBERS

APP. A

0123456789 In contrast, binary numbers do not use these ten digits. They are all constructed exclusively from the two binary digits 01 Octal numbers are built up from the eight octal digits 01234567 For hexadecimal numbers, 16 digits are needed. Thus six new symbols are required. It is conventional to use the uppercase letters A through F for the six digits following 9. Hexadecimal numbers are then built up from the digits 0123456789ABCDEF The expression ‘‘binary digit’’ meaning a 1 or a 0 is usually referred to as a bit. Figure A-2 shows the decimal number 2001 expressed in binary, octal, decimal, and hexadecimal form. The number 7B9 is obviously hexadecimal, because the symbol B can only occur in hexadecimal numbers. However, the number 111 might be in any of the four number systems discussed. To avoid ambiguity, people use a subscript of 2, 8, 10, or 16 to indicate the radix when it is not obvious from the context. Binary

1

1

1

1

1

0

1

0

0

0

1

1 × 210 + 1 × 29 + 1 × 28 + 1 × 27 + 1 × 26 + 0 × 25 + 1 × 24 + 0 × 23 + 0 × 22 + 0 × 21 + 1 × 20 +0 +0 +0 +0 + 16 +1 1024 + 512 + 256 + 128 + 64 Octal

3

7

2

1

3 × 83 + 7 × 82 + 2 × 81 + 1 × 80 1536 + 448 + 16 + 1 Decimal

2

0

0

1

2 × 103 + 0 × 102 + 0 × 101 + 1 × 100 2000 + 0 +0 +1 Hexadecimal

7

D 2

.

1 × 161

7 × 16 + 13 1792 + 208

+1 +1

× 160

Figure A-2. The number 2001 in binary, octal, and hexadecimal.

As an example of binary, octal, decimal, and hexadecimal notation, consider Fig. A-3, which shows a collection of nonnegative integers expressed in each of these four different systems. Perhaps some archaeologist thousands of years from now will discover this table and regard it as the Rosetta Stone to late twentieth century and early twenty-first century number systems.

SEC. A.3

CONVERSION FROM ONE RADIX TO ANOTHER

683

######################################### ! Octal ! Hex ! Decimal ! Binary #!! ######################################## ! ! ! ! 0 0 0 ! 0 ! #! ######################################## ! ! 1 ! 1 ! 1 ! 1 ! #!! ######################################## ! ! ! ! 2 10 2 2 ! #! ######################################## ! ! ! 3 ! 11 ! 3 ! 3 ! #!! ######################################## ! ! ! ! 4 100 3 3 ! #! ######################################## ! ! ! 5 ! 101 ! 5 ! 5 ! #!! ######################################## ! ! ! ! 6 110 6 6 ! #! ######################################## ! ! ! 7 ! 111 ! 7 ! 7 ! #!! ######################################## ! ! ! ! 8 1000 10 8 ! #! ######################################## ! ! ! 9 ! 1001 ! 11 ! 9 ! #!! ######################################## ! ! ! ! 10 1010 12 A #! ######################################## ! ! ! ! 11 ! 1011 ! 13 ! B ! #!! ######################################## ! ! ! ! 12 ! 1100 ! 14 ! C ! #! ######################################## 13 ! 1101 ! 15 ! D ! #!! ######################################## ! ! ! ! 14 1110 16 E ! #! ######################################## ! ! ! 15 ! 1111 ! 17 ! F ! #!! ######################################## ! ! ! ! 16 10000 20 10 #! ######################################## ! ! ! ! 20 ! 10100 ! 24 ! 14 ! #!! ######################################## ! ! ! ! 30 ! 11110 ! 36 ! 1E ! #! ######################################## 40 ! 101000 ! 50 ! 28 ! #!! ######################################## ! ! ! ! 50 110010 62 32 ! #! ######################################## ! ! ! 60 ! 111100 ! 74 ! 3C ! #!! ######################################## ! ! ! ! 70 1000110 106 46 ! #! ######################################## ! ! ! 80 ! 1010000 ! 120 ! 50 ! #!! ######################################## ! ! ! ! 90 1011010 132 5A #! ######################################## ! ! ! ! 100 ! 11001000 ! 144 ! 64 ! #!! ######################################## ! ! ! ! 1000 ! 1111101000 ! 1750 ! 3E8 ! #! ######################################## !!######################################### 2989 !! 101110101101 !! 5655 !! BAD !!

Figure A-3. Decimal numbers and their binary, octal, and hexadecimal equivalents.

A.3 CONVERSION FROM ONE RADIX TO ANOTHER Conversion between octal or hexadecimal numbers and binary numbers is easy. To convert a binary number to octal, divide it into groups of 3 bits, with the 3 bits immediately to the left (or right) of the decimal point (often called a binary point) forming one group, the 3 bits immediately to their left, another group, and so on. Each group of 3 bits can be directly converted to a single octal digit, 0 to 7, according to the conversion given in the first lines of Fig. A-3. It may be necessary to add one or two leading or trailing zeros to fill out a group to 3 full bits. Conversion from octal to binary is equally trivial. Each octal digit is simply replaced by the equivalent 3-bit binary number. Conversion from hexadecimal to

684

BINARY NUMBERS

APP. A

binary is essentially the same as octal-to-binary except that each hexadecimal digit corresponds to a group of 4 bits instead of 3 bits. Figure A-4 gives some examples of conversions. Example 1 Hexadecimal Binary Octal

. B 6 0 0 0 1 1 0 0 1 0 1 0 0 1 0 0 0. 1 0 1 1 0 1 1 0 0 5 0 . 5 1 4 1 4 5 1

9

4

8

Example 2 Hexadecimal Binary Octal

4 C . B 0 1 1 1 1 0 1 1 1 0 1 0 0 0 1 1. 1 0 1 1 1 1 0 0 0 1 0 0 4 5 0 7 3 . 5 7 4 6 7

B

A

3

Figure A-4. Examples of octal-to-binary and hexadecimal-to-binary conversion.

Conversion of decimal numbers to binary can be done in two different ways. The first method follows directly from the definition of binary numbers. The largest power of 2 smaller than the number is subtracted from the number. The process is then repeated on the difference. Once the number has been decomposed into powers of 2, the binary number can be assembled with 1s in the bit positions corresponding to powers of 2 used in the decomposition, and 0s elsewhere. The other method (for integers only) consists of dividing the number by 2. The quotient is written directly beneath the original number and the remainder, 0 or 1, is written next to the quotient. The quotient is then considered and the process repeated until the number 0 has been reached. The result of this process will be two columns of numbers, the quotients and the remainders. The binary number can now be read directly from the remainder column starting at the bottom. Figure A-5 gives an example of decimal-to-binary conversion. Binary integers can also be converted to decimal in two ways. One method consists of summing up the powers of 2 corresponding to the 1 bits in the number. For example, 10110 = 24 + 22 + 21 = 16 + 4 + 2 = 22 In the other method, the binary number is written vertically, one bit per line, with the leftmost bit on the bottom. The bottom line is called line 1, the one above it line 2, and so on. The decimal number will be built up in a parallel column next to the binary number. Begin by writing a 1 on line 1. The entry on line n consists of two times the entry on line n − 1 plus the bit on line n (either 0 or 1). The entry on the top line is the answer. Figure A-6 gives an example of this method of binary to decimal conversion.

SEC. A.3

CONVERSION FROM ONE RADIX TO ANOTHER Quotients

685

Remainders

1492 746

0

373

0

186

1

93

0

46

1

23

0

11

1

5

1

2

1

1

0

0

1

1 0 1 1 1 0 1 0 1 0 0 = 149210

Figure A-5. Conversion of the decimal number 1492 to binary by successive halving, starting at the top and working downward. For example, 93 divided by 2 yields a quotient of 46 and a remainder of 1, written on the line below it.

Decimal-to-octal and decimal-to-hexadecimal conversion can be accomplished either by first converting to binary and then to the desired system or by subtracting powers of 8 or 16.

A.4 NEGATIVE BINARY NUMBERS Four different systems for representing negative numbers have been used in digital computers at one time or another in history. The first one is called signed magnitude. In this system the leftmost bit is the sign bit (0 is + and 1 is −) and the remaining bits hold the absolute magnitude of the number. The second system, called one’s complement, also has a sign bit with 0 used for plus and 1 for minus. To negate a number, replace each 1 by a 0 and each 0 by a 1. This holds for the sign bit as well. One’s complement is obsolete. The third system, called two’s complement, also has a sign bit that is 0 for plus and 1 for minus. Negating a number is a two-step process. First, each 1 is

686 1

BINARY NUMBERS 0

1

1

1

0

1

1

0

1

1

APP. A

1 1 + 2 × 1499 = 2999

Result

1 + 2 × 749 = 1499 1 + 2 × 374 = 749 0 + 2 × 187 = 374 1 + 2 × 93 = 187 1 + 2 × 46 = 93 0 + 2 × 23 = 46 1 + 2 × 11 = 23 1 + 2 × 5 = 11 1+2×2=5 0+2×1=2 1+2×0=1

Start here

Figure A-6. Conversion of the binary number 101110110111 to decimal by successive doubling, starting at the bottom. Each line is formed by doubling the one below it and adding the corresponding bit. For example, 749 is twice 374 plus the 1 bit on the same line as 749.

replaced by a 0 and each 0 by a 1, just as in one’s complement. Second, 1 is added to the result. Binary addition is the same as decimal addition except that a carry is generated if the sum is greater than 1 rather than greater than 9. For example, converting 6 to two’s complement is done in two steps: 00000110 (+6) 11111001 (−6 in one’s complement) 11111010 (−6 in two’s complement) If a carry occurs from the leftmost bit, it is thrown away. The fourth system, which for m-bit numbers is called excess 2m − 1 , represents a number by storing it as the sum of itself and 2m − 1 . For example, for 8-bit numbers, m = 8, the system is called excess 128 and a number is stored as its true value plus 128. Therefore, −3 becomes −3 + 128 = 125, and −3 is represented by the 8-bit binary number for 125 (01111101). The numbers from −128 to +127 map onto 0 to 255, all of which are expressible as an 8-bit positive integer. Interestingly enough, this system is identical to two’s complement with the sign bit reversed. Figure A-7 gives examples of negative numbers in all four systems. Both signed magnitude and one’s complement have two representations for zero: a plus zero, and a minus zero. This situation is undesirable. The two’s complement system does not have this problem because the two’s complement of plus

SEC. A.4

NEGATIVE BINARY NUMBERS

687

#################################################################################### ! ! ! ! ! ! N N −N −N −N −N ! ! ! ! ! ! ! decimal ! binary ! signed mag. ! 1’s compl. ! 2’s compl. ! excess 128 ! !#################################################################################### !#################################################################################### 1 ! 00000001 ! 10000001 ! 11111110 ! 11111111 ! 01111111 ! ! ! ! ! ! ! ! 2 ! 00000010 ! 10000010 ! 11111101 ! 11111110 ! 01111110 ! !#################################################################################### ! 3 ! 00000011 ! 10000011 ! 11111100 ! 11111101 ! 01111101 ! !#################################################################################### ! ! ! ! ! ! 4 ! 00000100 ! 10000100 ! 11111011 ! 11111100 ! 01111100 ! !#################################################################################### ! 5 ! 00000101 ! 10000101 ! 11111010 ! 11111011 ! 01111011 ! !#################################################################################### ! ! ! ! ! ! !#################################################################################### 6 ! 00000110 ! 10000110 ! 11111001 ! 11111010 ! 01111010 ! ! ! ! ! ! ! ! 7 ! 00000111 ! 10000111 ! 11111000 ! 11111001 ! 01111001 ! !#################################################################################### !#################################################################################### 8 ! 00001000 ! 10001000 ! 11110111 ! 11111000 ! 01111000 ! ! ! ! ! ! ! ! 9 ! 00001001 ! 10001001 ! 11110110 ! 11110111 ! 01110111 ! !#################################################################################### ! 10 ! 00001010 ! 10001010 ! 11110101 ! 11110110 ! 01110110 ! !#################################################################################### ! ! ! ! ! ! 20 ! 00010100 ! 10010100 ! 11101011 ! 11101100 ! 01101100 ! !#################################################################################### ! ! ! ! ! ! ! 30 ! 00011110 ! 10011110 ! 11100001 ! 11100010 ! 01100010 ! !#################################################################################### !#################################################################################### 01011000 ! 11011000 ! 11010111 ! 10101000 ! 00101000 ! 40 ! ! ! ! ! ! ! ! 50 ! 00110010 ! 10110010 ! 11001101 ! 11001110 ! 01001110 ! !#################################################################################### ! 60 ! 00111100 ! 10111100 ! 11000011 ! 11000100 ! 01000100 ! !#################################################################################### ! ! ! ! ! ! 00111010 ! 10111010 ! 10111001 ! 11000110 ! 01000110 ! 70 ! !#################################################################################### ! 80 ! 01010000 ! 11010000 ! 10101111 ! 10110000 ! 00110000 ! !#################################################################################### ! ! ! ! ! ! 90 ! 01011010 ! 11011010 ! 10100101 ! 10100110 ! 00100110 ! !#################################################################################### ! ! ! ! ! ! ! 00011100 ! 10011100 ! 10011011 ! 11100100 ! 01100100 ! 100 ! !#################################################################################### !#################################################################################### 127 ! 01111111 ! 11111111 ! 10000000 ! 10000001 ! 00000001 ! ! ! ! ! ! ! ! 128 Nonexistent Nonexistent Nonexistent 10000000 00000000 !#################################################################################### ! ! ! ! ! ! !

Figure A-7. Negative 8-bit numbers in four systems.

zero is also plus zero. The two’s complement system does, however, have a different singularity. The bit pattern consisting of a 1 followed by all 0s is its own complement. The result is to make the range of positive and negative numbers unsymmetric; there is one negative number with no positive counterpart. The reason for these problems is not hard to find: we want an encoding system with two properties: 1. Only one representation for zero. 2. Exactly as many positive numbers as negative numbers. The problem is that any set of numbers with as many positive as negative numbers and only one zero has an odd number of members, whereas m bits allow an even number of bit patterns. There will always be either one bit pattern too many or one bit pattern too few, no matter what representation is chosen. This extra bit

688

BINARY NUMBERS

APP. A

pattern can be used for −0 or for a large negative number, or for something else, but no matter what it is used for it will always be a nuisance.

A.5 BINARY ARITHMETIC The addition table for binary numbers is given in Fig. A-8. Addend Augend Sum Carry

0 +0 "" 0 0

0 +1 "" 1 0

1 +0 "" 1 0

1 +1 "" 0 1

Figure A-8. The addition table in binary.

Two binary numbers can be added, starting at the rightmost bit and adding the corresponding bits in the addend and the augend. If a carry is generated, it is carried one position to the left, just as in decimal arithmetic. In one’s complement arithmetic, a carry generated by the addition of the leftmost bits is added to the rightmost bit. This process is called an end-around carry. In two’s complement arithmetic, a carry generated by the addition of the leftmost bits is merely thrown away. Examples of binary arithmetic are shown in Fig. A-9. Decimal

1's complement

2's complement

10 + (−3)

00001010 11111100

00001010 11111101

+7

1 00000110

1 00000111

carry 1

discarded

00000111

Figure A-9. Addition in one’s complement and two’s complement.

If the addend and the augend are of opposite signs, overflow error cannot occur. If they are of the same sign and the result is of the opposite sign, overflow error has occurred and the answer is wrong. In both one’s and two’s complement arithmetic, overflow occurs if and only if the carry into the sign bit differs from the carry out of the sign bit. Most computers preserve the carry out of the sign bit, but the carry into the sign bit is not visible from the answer. For this reason, a special overflow bit is usually provided.

APP. A

689

PROBLEMS

PROBLEMS 1. Convert the following numbers to binary: 1984, 4000, 8192. 2. What is 1001101001 (binary) in decimal? In octal? In hexadecimal? 3. Which of the following are valid hexadecimal numbers? BED, CAB, DEAD, DECADE, ACCEDED, BAG, DAD. 4. Express the decimal number 100 in all radices from 2 to 9. 5. How many different positive integers can be expressed in k digits using radix r numbers? 6. Most people can only count to 10 on their fingers; however, computer scientists can do better. If you regard each finger as one binary bit, with finger extended as 1 and finger touching palm as 0, how high can you count using both hands? With both hands and both feet? Now use both hands and both feet, with the big toe on your left foot as a sign bit for two’s complement numbers. What is the range of expressible numbers? 7. Perform the following calculations on 8-bit two’s complement numbers. 00101101 + 01101111 """"""""

11111111 + 11111111 """"""""

00000000 − """""""" 11111111

11110111 − """""""" 11110111

8. Repeat the calculation of the preceding problem but now in one’s complement. 9. Consider the following addition problems for 3-bit binary numbers in two’s complement. For each sum, state a. Whether the sign bit of the result is 1. b. Whether the low-order 3 bits are 0. c. Whether an overflow occurred. 000 + """ 001

000 + """ 111

100 + """ 111

111 + """ 110

100 + """ 100

10. Signed decimal numbers consisting of n digits can be represented in n + 1 digits without a sign. Positive numbers have 0 as the leftmost digit. Negative numbers are formed by subtracting each digit from 9. Thus the negative of 014725 is 985274. Such numbers are called nine’s complement numbers and are analogous to one’s complement binary numbers. Express the following as three-digit nine’s complement numbers: 6, −2, 100, −14, −1, 0. 11. Determine the rule for addition of nine’s complement numbers and then perform the following additions. 0001 + 9999 """"

0001 + 9998 """"

9997 + 9996 """"

9241 + 0802 """"

12. Ten’s complement is analogous to two’s complement. A ten’s complement negative number is formed by adding 1 to the corresponding nine’s complement number, ignoring the carry. What is the rule for ten’s complement addition?

690

BINARY NUMBERS

APP. A

13. Construct the multiplication tables for radix 3 numbers. 14. Multiply 0111 and 0011 in binary. 15. Write a program that takes in a signed decimal number as an ASCII string and prints out its representation in two’s complement in binary, octal, and hexadecimal. 16. Write a program that takes in two 32-character ASCII strings containing 0s and 1s, each representing a two’s complement 32-bit binary number. The program should print their sum as a 32-character ASCII string of 0s and 1s.

B FLOATING-POINT NUMBERS

In many calculations the range of numbers used is very large. For example, a calculation in astronomy might involve the mass of the electron, 9 × 10−28 grams, and the mass of the sun, 2 × 1033 grams, a range exceeding 1060 . These numbers could be represented by 0000000000000000000000000000000000.0000000000000000000000000009 2000000000000000000000000000000000.0000000000000000000000000000 and all calculations could be carried out keeping 34 digits to the left of the decimal point and 28 places to the right of it. Doing so would allow 62 significant digits in the results. On a binary computer, multiple-precision arithmetic could be used to provide enough significance. However, the mass of the sun is not even known accurately to five significant digits, let alone 62. In fact few measurements of any kind can (or need) be made accurately to 62 significant digits. Although it would be possible to keep all intermediate results to 62 significant digits and then throw away 50 or 60 of them before printing the final results, doing this is wasteful of both CPU time and memory. What is needed is a system for representing numbers in which the range of expressible numbers is independent of the number of significant digits. In this appendix, such a system will be discussed. It is based on the scientific notation commonly used in physics, chemistry, and engineering.

691

692

FLOATING-POINT NUMBERS

APP. B

B.1 PRINCIPLES OF FLOATING POINT One way of separating the range from the precision is to express numbers in the familiar scientific notation n = f × 10e where f is called the fraction, or mantissa, and e is a positive or negative integer called the exponent. The computer version of this notation is called floating point. Some examples of numbers expressed in this form are 3.14 = 0.314 × 101 = 3.14 × 100 0.000001 = 0.1 × 10−5 = 1.0 × 10−6 1941 = 0.1941 × 104 = 1.941 × 103 The range is effectively determined by the number of digits in the exponent and the precision is determined by the number of digits in the fraction. Because there is more than one way to represent a given number, one form is usually chosen as the standard. In order to investigate the properties of this method of representing numbers, consider a representation, R, with a signed three-digit fraction in the range 0.1 ≤ ! f ! < 1 or zero and a signed two-digit exponent. These numbers range in magnitude from +0.100 × 10−99 to +0.999 × 10+99 , a span of nearly 199 orders of magnitude, yet only five digits and two signs are needed to store a number. Floating-point numbers can be used to model the real-number system of mathematics, although there are some important differences. Figure B-1 gives a grossly exaggerated schematic of the real number line. The real line is divided up into seven regions: 1. Large negative numbers less than −0.999 × 1099 . 2. Negative numbers between −0.999 × 1099 and −0.100 × 10−99 . 3. Small negative numbers with magnitudes less than 0.100 × 10−99 . 4. Zero. 5. Small positive numbers with magnitudes less than 0.100 × 10−99 . 6. Positive numbers between 0.100 × 10−99 and 0.999 × 1099 . 7. Large positive numbers greater than 0.999 × 1099 . One major difference between the set of numbers representable with three fraction and two exponent digits and the real numbers is that the former cannot be used to express any numbers in regions 1, 3, 5, or 7. If the result of an arithmetic operation yields a number in regions 1 or 7—for example, 1060 × 1060 = 10120 — overflow error will occur and the answer will be incorrect. The reason is due to the finite nature of the representation for numbers and is unavoidable. Similarly,

SEC. B.1

693

PRINCIPLES OF FLOATING POINT 3 Negative underflow 2 Expressible negative numbers

1 Negative overflow —1099

4 Zero

—10—100 0

5 Positive underflow 7 Positive overflow

6 Expressible positive numbers 10—100

1099

Figure B-1. The real number line can be divided into seven regions.

a result in regions 3 or 5 cannot be expressed either. This situation is called underflow error. Underflow error is less serious than overflow error, because 0 is often a satisfactory approximation to numbers in regions 3 and 5. A bank balance of 10−102 dollars is hardly better than a bank balance of 0. Another important difference between floating-point numbers and real numbers is their density. Between any two real numbers, x and y, is another real number, no matter how close x is to y. This property comes from the fact that for any distinct real numbers, x and y, z = (x + y)/2 is a real number between them. The real numbers form a continuum. Floating-point numbers, in contrast, do not form a continuum. Exactly 179,100 positive numbers can be expressed in the five-digit, two-sign system used above, 179,100 negative numbers, and 0 (which can be expressed in many ways), for a total of 358,201 numbers. Of the infinite number of real numbers between −10+100 and +0.999 × 1099 , only 358,201 of them can be specified by this notation. They are symbolized by the dots in Fig. B-1. It is quite possible for the result of a calculation to be one of the other numbers, even though it is in region 2 or 6. For example, +0.100 × 103 divided by 3 cannot be expressed exactly in our system of representation. If the result of a calculation cannot be expressed in the number representation being used, the obvious thing to do is to use the nearest number that can be expressed. This process is called rounding. The spacing between adjacent expressible numbers is not constant throughout region 2 or 6. The separation between +0.998 × 1099 and +0.999 × 1099 is vastly more than the separation between +0.998 × 100 and +0.999 × 100 . However, when the separation between a number and its successor is expressed as a percentage of that number, there is no systematic variation throughout region 2 or 6. In other words, the relative error introduced by rounding is approximately the same for small numbers as large numbers. Although the preceding discussion was in terms of a representation system with a three-digit fraction and a two-digit exponent, the conclusions drawn are valid for other representation systems as well. Changing the number of digits in the fraction or exponent merely shifts the boundaries of regions 2 and 6 and changes the number of expressible points in them. Increasing the number of digits in the fraction increases the density of points and therefore improves the accuracy

694

FLOATING-POINT NUMBERS

APP. B

of approximations. Increasing the number of digits in the exponent increases the size of regions 2 and 6 by shrinking regions 1, 3, 5, and 7. Figure B-2 shows the approximate boundaries of region 6 for floating-point decimal numbers for various sizes of fraction and exponent. """""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! Digits in fraction ! Digits in exponent ! Lower bound ! Upper bound ! !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −12 9 10 3 1 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −102 99 ! ! ! ! ! 10 3 2 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −1002 999 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! 10 3 3 10 ! ! ! ! ! −10002 9999 10 3 4 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −13 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! 109 4 1 10 ! ! ! ! ! −103 99 10 4 2 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −1003 999 ! ! ! ! ! 10 4 3 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −10003 9999 10 4 4 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −14 9 ! ! ! ! ! 10 5 1 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −104 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! 1099 5 2 10 ! ! ! ! ! −1004 999 10 5 3 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −10004 9999 ! ! ! ! ! 10 5 4 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −1009 999 10 10 3 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! −1019 999 ! ! ! ! ! 10 20 3 10 !"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! Figure B-2. The approximate lower and upper bounds of expressible (unnormalized) floating-point decimal numbers.

A variation of this representation is used in computers. For efficiency, exponentiation is to base 2, 4, 8, or 16 rather than 10, in which case the fraction consists of a string of binary, base-4, octal, or hexadecimal digits. If the leftmost of these digits is zero, all the digits can be shifted one place to the left and the exponent decreased by 1, without changing the value of the number (barring underflow). A fraction with a nonzero leftmost digit is said to be normalized. Normalized numbers are generally preferable to unnormalized numbers, because there is only one normalized form, whereas there are many unnormalized forms. Examples of normalized floating-point numbers are given in Fig. B-3 for two bases of exponentiation. In these examples a 16-bit fraction (including sign bit) and a 7-bit exponent using excess 64 notation are shown. The radix point is to the left of the leftmost fraction bit—that is, to the right of the exponent.

B.2 IEEE FLOATING-POINT STANDARD 754 Until about 1980, each computer manufacturer had its own floating-point format. Needless to say, all were different. Worse yet, some of them actually did arithmetic incorrectly because floating-point arithmetic has some subtleties not obvious to the average hardware designer.

SEC. B.2

IEEE FLOATING-POINT STANDARD 754

695

Example 1: Exponentiation to the base 2 2–2 –1

2

Unnormalized:

0 1010100

2–4 2

–3

2–6 2

–5

2–8 –7

2

2–10 –9

2–12 –11

2

2

2–14 –13

2

2

2–16 –15

.0

20 –12 –13 –15 0 0 0 0 0 0 0 0 0 0 1 1 0 1 1 = 2 (1 × 2 + 1 × 2 + 1 × 2

.1

1 0 1 1 0 0 0 0 0 0 0 0 0 0 0 = 29 (1 × 2–1+ 1 × 2–2+ 1 × 2–4

+ 1 × 2–16) = 432 Sign Excess 64 Fraction is 1 × 2–12+ 1 × 2–13 + exponent is +1 × 2–15+ 1 × 2–16 84 – 64 = 20 To normalize, shift the fraction left 11 bits and subtract 11 from the exponent. Normalized:

0 1001001

+ 1 × 2–5) = 432

Fraction is 1 × 2–1 + 1 × 2–2 +1 × 2–4 + 1 × 2–5

Sign Excess 64 + exponent is 73 – 64 = 9

Example 2: Exponentiation to the base 16

Unnormalized:

0 1000101

.

16–1

16–2

0 0 00

0 0 00

Sign Excess 64 + exponent is 69 – 64 = 5

Fraction is 1 × 16

16–3

16–4 1 0 1 1 = 165 (1 × 16–3+ B × 16–4) = 432

0 0 01 –3

+ B × 16

–4

To normalize, shift the fraction left 2 hexadecimal digits, and subtract 2 from the exponent. Normalized:

0 1000011 Sign Excess 64 + exponent is 67 – 64 = 3

.

0001

1011 Fraction is 1 × 16

0 0 0 0 = 163 (1 × 16–1+ B × 16–2) = 432

0000 –1

+ B × 16

–2

Figure B-3. Examples of normalized floating-point numbers.

To rectify this situation, in the late 1970s IEEE set up a committee to standardize floating-point arithmetic. The goal was not only to permit floating-point data to be exchanged among different computers but also to provide hardware designers with a model known to be correct. The resulting work led to IEEE Standard 754 (IEEE, 1985). Most CPUs these days (including the Intel, SPARC, and JVM ones studied in this book) have floating-point instructions that conform to the IEEE floating-point standard. Unlike many standards, which tend to be wishy-washy compromises that please no one, this one is not bad, in large part because it was primarily the work of one person, Berkeley math professor William Kahan. The standard will be described in the remainder of this section. The standard defines three formats: single precision (32 bits), double precision (64 bits), and extended precision (80 bits). The extended-precision format is intended to reduce roundoff errors. It is used primarily inside floating-point arithmetic units, so we will not discuss it further. Both the single- and doubleprecision formats use radix 2 for fractions and excess notation for exponents. The formats are shown in Fig. B-4. Both formats start with a sign bit for the number as a whole, 0 being positive and 1 being negative. Next comes the exponent, using excess 127 for single

696

FLOATING-POINT NUMBERS Bits 1

8

APP. B

23 Fraction

Sign

Exponent (a)

Bits 1

11

52

Exponent

Fraction

Sign (b)

Figure B-4. IEEE floating-point formats. (a) Single precision. (b) Double precision.

precision and excess 1023 for double precision. The minimum (0) and maximum (255 and 2047) exponents are not used for normalized numbers; they have special uses described below. Finally, we have the fractions, 23 and 52 bits, respectively. A normalized fraction begins with a binary point, followed by a 1 bit, and then the rest of the fraction. Following a practice started on the PDP-11, the authors of the standard realized that the leading 1 bit in the fraction does not have to be stored, since it can just be assumed to be present. Consequently, the standard defines the fraction in a slightly different way than usual. It consists of an implied 1 bit, an implied binary point, and then either 23 or 52 arbitrary bits. If all 23 or 52 fraction bits are 0s, the fraction has the numerical value 1.0; if all of them are 1s, the fraction is numerically slightly less than 2.0. To avoid confusion with a conventional fraction, the combination of the implied 1, the implied binary point, and the 23 or 52 explicit bits is called a significand instead of a fraction or mantissa. All normalized numbers have a significand, s, in the range 1 ≤ s < 2. The numerical characteristics of the IEEE floating-point numbers are given in Fig. B-5. As examples, consider the numbers 0.5, 1, and 1.5 in normalized single-precision format. These are represented in hexadecimal as 3F000000, 3F800000, and 3FC00000, respectively. One of the traditional problems with floating-point numbers is how to deal with underflow, overflow, and uninitialized numbers. The IEEE standard deals with these problems explicitly, borrowing its approach in part from the CDC 6600. In addition to normalized numbers, the standard has four other numerical types, described below and shown in Fig. B-6. A problem arises when the result of a calculation has a magnitude smaller than the smallest normalized floating-point number that can be represented in this system. Previously, most hardware took one of two approaches: just set the result to zero and continue, or cause a floating-point underflow trap. Neither of these is

SEC. B.2

IEEE FLOATING-POINT STANDARD 754

697

" """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! Single precision ! Double precision ! Item "! """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! in sign 1 1 ! ! ! "! Bits """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! Bits in exponent ! ! ! 8 11 "! """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! ! ! in fraction 23 52 "! Bits """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! Bits, total 32 64 "! """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! ! ! system Excess 127 Excess 1023 "! Exponent """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! ! range −126 to +127 −1022 to +1023 !"Exponent ! ! ! """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" −126 −1022 ! Smallest normalized number ! ! ! 2 2 "! """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! 128 1024 approx. 2 normalized number approx. 2 ! ! ! "! Largest """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! Decimal range ! approx. 10−38 to 1038 ! approx. 10−308 to 10308 ! "! """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" ! ! ! −45 !!"Smallest !! !! approx. 10−324 denormalized number!! approx. 10 """"""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""""" Figure B-5. Characteristics of IEEE floating-point numbers.

Normalized ±

0 < Exp < Max

Any bit pattern

Denormalized ±

0

Any nonzero bit pattern

Zero ±

0

0

Infinity ±

1 1 1…1

0

Not a number ±

1 1 1…1

Any nonzero bit pattern

Sign bit

Figure B-6. IEEE numerical types.

really satisfactory, so IEEE invented denormalized numbers. These numbers have an exponent of 0 and a fraction given by the following 23 or 52 bits. The implicit 1 bit to the left of the binary point now becomes a 0. Denormalized numbers can be distinguished from normalized ones because the latter are not permitted to have an exponent of 0. The smallest normalized single precision number has a 1 as exponent and 0 as fraction, and represents 1.0 × 2−126 . The largest denormalized number has a 0 as exponent and all 1s in the fraction, and represents about 0.9999999 × 2−126 , which is almost the same thing. One thing to note however, is that this number has only 23 bits of significance, versus 24 for all normalized numbers. As calculations further decrease this result, the exponent stays put at 0, but the first few bits of the fraction become zeros, reducing both the value and the number of significant bits in the fraction. The smallest nonzero denormalized

698

FLOATING-POINT NUMBERS

APP. B

number consists of a 1 in the rightmost bit, with the rest being 0. The exponent represents 2−126 and the fraction represents 2−23 so the value is 2−149 . This scheme provides for a graceful underflow by giving up significance instead of jumping to 0 when the result cannot be expressed as a normalized number. Two zeros are present in this scheme, positive and negative, determined by the sign bit. Both have an exponent of 0 and a fraction of 0. Here too, the bit to the left of the binary point is implicitly 0 rather than 1. Overflow cannot be handled gracefully. There are no bit combinations left. Instead, a special representation is provided for infinity, consisting of an exponent with all 1s (not allowed for normalized numbers), and a fraction of 0. This number can be used as an operand and behaves according to the usual mathematical rules for infinity. For example infinity plus anything is infinity, and any finite number divided by infinity is zero. Similarly, any finite number divided by zero yields infinity. What about infinity divided by infinity? The result is undefined. To handle this case, another special format is provided, called NaN (Not a Number). It too, can be used as an operand with predictable results.

PROBLEMS 1. Convert the following numbers to IEEE single-precision format. Give the results as eight hexadecimal digits. a. 9 b. 5/32 c. −5/32 d. 6.125 2. Convert the following IEEE single-precision floating-point numbers from hex to decimal: a. 42E48000H b. 3F880000H c. 00800000H d. C7F00000H 3. The format of single-precision floating-point numbers on the 370 has a 7-bit exponent in the excess 64 system, and a fraction containing 24 bits plus a sign bit, with the binary point at the left end of the fraction. The radix for exponentiation is 16. The order of the fields is sign bit, exponent, fraction. Express the number 7/64 as a normalized number in this system in hex. 4. The following binary floating-point numbers consist of a sign bit, an excess 64, radix 2 exponent, and a 16-bit fraction. Normalize them. a. 0 1000000 0001010100000001

APP. B

PROBLEMS

699

b. 0 0111111 0000001111111111 c. 0 1000011 1000000000000000 5. To add two floating-point numbers, you must adjust the exponents (by shifting the fraction) to make them the same. Then you can add the fractions and normalize the result, if need be. Add the single-precision IEEE numbers 3EE00000H and 3D800000H and express the normalized result in hexadecimal. 6. The Tightwad Computer Company has decided to come out with a machine having 16-bit floating-point numbers. The Model 0.001 has a floating-point format with a sign bit, 7-bit, excess 64 exponent, and 8-bit fraction. The Model 0.002 has a sign bit, 5-bit, excess 16 exponent, and 10-bit fraction. Both use radix 2 exponentiation. What are the smallest and largest positive normalized numbers on both models? About how many decimal digits of precision does each have? Would you buy either one? 7. There is one situation in which an operation on two floating-point numbers can cause a drastic reduction in the number of significant bits in the result. What is it? 8. Some floating-point chips have a square root instruction built in. A possible algorithm is an iterative one (e.g., Newton-Raphson). Iterative algorithms need an initial approximation and then steadily improve it. How can one obtain a fast approximate square root of a floating-point number? 9. Write a procedure to add two IEEE single-precision floating-point numbers. Each number is represented by a 32-element Boolean array. 10. Write a procedure to add two single-precision floating-point numbers that use radix 16 for the exponent and radix 2 for the fraction but do not have an implied 1 bit to the left of the binary point. A normalized number has 0001, 0010, ..., 1111 as the leftmost 4 bits of the fraction, but not 0000. A number is normalized by shifting the fraction left 4 bits and subtracting 1 from the exponent.

This page intentionally left blank

ec

C ASSEMBLY LANGUAGE PROGRAMMING

Evert Wattel Vrije Universiteit Amsterdam, The Netherlands

Every computer has an ISA (Instruction Set Architecture), which is a set of registers, instructions, and other features visible to its low-level programmers. This ISA is commonly referred to as machine language, although the term is not entirely accurate. A program at this level of abstraction is a long list of binary numbers, one per instruction, telling which instructions to execute and what their operands are. Programming with binary numbers is very difficult to do, so all machines have an assembly language, a symbolic representation of the instruction set architecture, with symbolic names like, ADD, SUB, and MUL, instead of binary numbers. This appendix is a tutorial on assembly language programming for one specific machine, the Intel 8088, which was used in the original IBM PC and was the base from which the modern Pentium grew. The appendix also covers the use of some tools that can be downloaded to help learn about assembly language programming. The purpose of this appendix is not to turn out polished assembly language programmers, but to help the reader learn about computer architecture through 701

702

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

hands-on experience. For this reason, a simple machine—the Intel 8088— has been chosen as the running example. While 8088s are rarely encountered any more, every Pentium is capable of executing 8088 programs, so the lessons learned here are still applicable to modern machines. Furthermore, most of the Pentium’s core instructions are the same as the 8088’s, only using 32-bit registers instead of 16-bit registers. Thus, this appendix can also be seen as a gentle introduction to Pentium assembly language programming. In order to program any machine in assembly language, the programmer must have a detailed knowledge of the machine’s instruction set architecture. Accordingly, Sections C.1 through C.4 of this appendix are devoted to the architecture of the 8088, its memory organization, addressing modes, and instructions. Section C.5 discusses the assembler, which is used in this appendix and which is available for free, as described later. The notation used in this appendix is the one used by this assembler. Other assemblers use different notations, so readers already familiar with 8088 assembly programming should be alert for differences. Section C.6 discusses an interpreter/tracer/debugger tool, which can be downloaded to help the beginner programmer get programs debugged. Section C.7 describes the installation of the tools, and how to get started. Section C.8 contain programs, examples, exercises and solutions. Section C.9 discusses implementation issues, bugs and limitations of the material.

C.1 OVERVIEW We will start our tour of assembly language programming with a few words on assembly language and then give a small example to illustrate it.

C.1.1 Assembly Language Every assembler uses mnemonics, that is, short words such as ADD, SUB, and for machine instructions such as add, subtract, and multiply, to make them easy to remember. In addition, assemblers allow the use of symbolic names for constants and labels to indicate instruction and memory addresses. Also, most assemblers support some number of pseudoinstructions, which do not translate into ISA instructions, but which are commands to the assembler to guide the assembly process. When a program in assembly language is fed to a program called an assembler, the assembler converts the program into a binary program suitable for actual execution. This program can then be run on the actual hardware. However, when beginners start to program in assembly language, they often make errors and the binary program just stops, without any clue as to what went wrong. To make life easier for beginners, it is sometimes possible to run the binary program not on the actual hardware, but on a simulator, which executes one MUL

SEC. C.1

703

OVERVIEW

instruction at a time and gives a detailed display of what it is doing. In this way, debugging is much easier. Programs running on a simulator run very slowly, of course, but when the goal is to learn assembly language program, rather than run a production job, this loss of speed is not important. This appendix is based on a toolkit that includes such a simulator, called the interpreter or tracer, as it interprets and traces the execution of the binary program step by step as it runs. The terms ‘‘simulator,’’ ‘‘interpreter,’’ and ‘‘tracer’’ will be used interchangeably throughout this appendix. Usually, when we are talking about just executing a program, we will speak of the ‘‘interpreter’’ and when we are talking about using it as a debugging tool, we will call it the ‘‘tracer,’’ but it is the same program.

C.1.2 A Small Assembly Language Program To make some of these abstract ideas a bit more concrete, consider the program and tracer image of Fig. C-1. An image of the tracer screen is given in Fig. C-1. Fig. C-1(a) shows a simple assembly language program for the 8088. The numbers following the exclamation marks are the source line numbers, to make it easier to refer to parts of the program. A copy of this program can be found in the accompanying material, in the directory examples in the source file HlloWrld.s. This assembly program, like all assembly programs discussed in this appendix, has the suffix .s, which indicates that it is an assembly language source program. The tracer screen, shown in Fig. C-1(b), contains seven windows, each containing different information about the state of the binary program being executed. _EXI T = 1 _WRITE = 4 _STDOUT = 1 .SECT .TEXT start: MOV CX,de-hw PUSH CX PUSH hw PUSH _STDOUT PUSH _WRITE SYS ADD SP, 8 SUB CX,AX PUSH CX PUSH _EXIT SYS .SECT .DATA hw: . ASCII "Hello World\n " de: .BYTE 0

(a)

! 1 ! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13 ! 14 ! 15 ! 16 ! 17 ! 18 ! 19 ! 20

C S : 00 D S = S S = E S : 002 AH:00 A L : 0 c AX: 12 BH:00 BL:00 B X : 0 CH:00 C L : 0 c C X : 12 DH:00 D L : 0 0 DX: 0 S P : 7 f d 8 SF O D S Z C => 0004 B P : 0000 CC - > p - 0001 S I : 0000 IP:000c:PC 0000 D I : 0000 start + 7 000c E I

=>

MOV PUSH PUSH PUSH PUSH SYS ADD SUB PUSH

CX,de-hw CX HW _STDOUT _WRITE SP , 8 CX,AX CX

! ! ! ! ! ! ! ! !

6 7 8 9 10 11 12 13 14

hw

> H e l l o World\n hw + 0 = 0 0 0 0 : 48 65 6c 6c 6f 20 57 6f H e l l o

World

(b)

Figure C-1. (a) An assembly language program. (b) The corresponding tracer display.

25928

704

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

Let us now briefly examine the seven windows of Fig. C-1(b). On the top are three windows, two larger ones and a smaller one in the middle. The top left window shows the contents of the processor, consisting of the current values of the segment registers, CS, DS, SS, and ES, the arithmetic registers, AH, AL, AX, and others. The middle window in the top row contains the stack, an area of memory used for temporary values. The right-hand window in the top row contains a fragment of the assembly language program, with the arrow showing which instruction is currently being executed. As the program runs, the current instruction changes and the arrow moves to point to it. The strength of the tracer is that by hitting the return key (labeled Enter on PC keyboards), one instruction is executed and all the windows are updated, making it possible to run the program in slow motion. Below the left window is a window that contains the subroutine call stack, here empty. Below it are commands to the tracer itself. To the left of these two windows is a window for input, output, and error messages. Below these windows is a window that shows a portion of memory. These windows will be discussed in more detail later, but the basic idea should be clear: the tracer shows the source program, the machine registers, and quite a bit of information about the state of the program being executed. As each instruction is executed the information is updated, allowing the user to see in great detail what the program is doing.

C.2 THE 8088 PROCESSOR Every processor, including the 8088, has an internal state, where it keeps certain crucial information. For this purpose, the processor has a set of registers where this information can be stored and processed. Probably the most important of these is the PC (program counter), which contains the memory location, that is, the address, of the next instruction to be executed. This register is also called IP (Instruction Pointer). This instruction is located in a part of the main memory, called the code segment. The main memory on the 8088 may be up to slightly more the 1 MB in size, but the current code segment is only 64 KB. The CS register in Fig. C-1 tells where the 64-KB code segment begins within the 1MB memory. A new code segment can be activated by simply changing the value of the CS register. Similarly, there is also a 64-KB data segment, which tells where the data begins. In Fig. C-1 its origin is given by the DS register, which can also be changed as needed to access data outside the current data segment. The CS and DS registers are needed because the 8088 has 16-bit registers, so they cannot directly hold the 20-bit addresses needed to reference the entire 1-MB memory. This is why the code and data segment registers were introduced.

SEC. C.2

THE 8088 PROCESSOR

705

The other registers contain data or pointers to data in the main memory. In assembly language programs, these registers can be directly accessed. Apart from these registers, the processor also contains all the necessary equipment to perform the instructions, but these parts are available to the programmer only through the instructions.

C.2.1 The Processor Cycle The operation of the 8088 (and all other computers) consists of executing instructions, one after another. The execution of a single instruction can be broken down into the following steps: 1. Fetch the instruction from memory from the code segment using PC. 2. Increment the program counter. 3. Decode the fetched instruction. 4. Fetch the necessary data from memory and/or processor registers. 5. Perform the instruction. 6. Store the results of the instruction in memory and/or registers. 7. Go back to step 1 to start the next instruction. The execution of an instruction is somewhat like running a very small program. In fact, some machines really do have a little program, called a microprogram, to execute their instructions. Microprograms are described in detail in Chap. 4. From the point of view of an assembly programmer, the 8088 has a set of 14 registers. These registers are in some sense the scratch pad where the instructions operate and are in constant use, although the results stored in them are very volatile. Figure C-2 gives an overview of these 14 registers. It is clear that this figure and the register window of the tracer of Fig. C-1 are very similar because they represent the same information. The 8088 registers are 16 bits wide. No two registers are completely functionally equivalent, but some of them share certain features, so they are subdivided into groups in Fig. C-2. We will now discuss the different groups.

C.2.2 The General Registers The registers in the first group, AX, BX, CX, and DX are the general registers. The first register of this group, AX, is called the accumulator register. It is used to collect results of computations and is the target of many of the instructions. Although each of the registers can perform a host of tasks, in some instructions this AX is the implied destination, for example, in multiplication.

706

ASSEMBLY LANGUAGE PROGRAMMING

General

APP. C

Segment registers

registers

AX

AH

AL

CS

Code segment

BX

BH

BL

DS

Data segment

CX

CH

CL

SS

Stack segment

DX

DH

DL

ES

E x t r a segment

15

8 7

0

15

P o i n t e r and index SP

Stack p o i n t e r

BP

Base p o i n t e r

SI

Source index

Condition SF 15

Destination 15

index

codes

OD I T S Z A S t a t u s flags Instruction

P

C 0

15

CC

pointer

Program counter

IP DI

0

PC 0

0

Figure C-2. The 8088 registers.

The second register of this group is BX, the base register. For many purposes can be used in the same way as AX, but it has one power AX does not have. It is possible to put a memory address in BX and then execute an instruction whose operand comes from the memory address contained in BX. In other words, BX can hold a pointer to memory, AX cannot. To show this, we compare two instructions. First we have BX

MOV AX,BX

which copies to AX the contents of BX. Second we have MOV AX,(BX)

which copies to AX the contents of the memory word whose address is contained in BX. In the first example, BX contains the source operand; in the second one it

SEC. C.2

THE 8088 PROCESSOR

707

points to the source operand. In both of these examples, note that the MOV instruction has a source and a destination operand, and that the destination is written before the source. The next general register is CX, the counter register. Besides fulfilling many other tasks, this register is specifically used to contain counters for loops. It is automatically decremented in the LOOP instruction, and loops are usually terminated when CX reaches zero. The fourth register of the general group is DX, the data register. It is used together with AX in double word length (i.e., 32-bit) instructions. In this case, DX contains the high-order 16 bits and AX contains the low-order 16 bits. Usually, 32-bit integers are indicated by the term long. The term double is usually reserved for 64-bit floating point values, although some people use ‘‘double’’ for 32-bit integers. In this tutorial, there will be no confusion because we will not discuss floating-point numbers at all. All of these general registers can be regarded either as a 16-bit register or as a pair of 8-bit registers. In this way, the 8088 has precisely eight different 8-bit registers, which can be used in byte and character instructions. None of the other registers can be split into 8-bit halves. Some instructions use an entire register, such as AX, but other instructions use only half of a register, such as AL or AH. In general, instructions doing arithmetic use the full 16-bit registers, but instructions dealing with characters usually use the 8-bit registers. It is important, however, to realize that AL and AH are just names for both halves of AX. When AX is loaded with a new value, both AL and AH are changed to the lower and upper halves of the 16-bit number put in AX, respectively. To see how AX, AH, and AL interact, consider the instruction MOV AX,258

which loads the AX register with the decimal value 258. After this instruction, the byte register AH contains the value 1, and the byte register AL contains the number 2. If this instruction is followed by the byte add instruction ADDB AH,AL

then the byte register AH is incremented by the value in AL (2) so that it now contains 3. The effect on the register AX of this action is that its value is now 770, which is equivalent to 00000011 00000010 in binary notation or 0x03 0x02 in hexadecimal notation. The eight byte-wide registers are almost interchangeable, with the exception that AL always contains one of the operands in the MULB instruction, and is the implied destination of this operation, together with AH. DIVB also uses the AH : AL pair for the dividend. The lower byte of the counter register CL can be used to hold the number of cycles in shift and rotate instruction. Section C.8, example 2, shows some of the properties of the general registers by means of a discussion of the program GenReg.s.

708

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

C.2.3 Pointer Registers The second group of registers consists of the pointer and index registers. The most important register of this group is the stack pointer, which is denoted by SP. Stacks are important in most programming languages. The stack is a segment of memory that holds certain context information about the running program. Usually, when a procedure is called, part of the stack is reserved for holding the procedure’s local variables, the address to return to when the procedure has finished, and other control information. The portion of the stack relating to a procedure is called its stack frame . When a called procedure calls another procedure, an additional stack frame is allocated, usually just below the current one. Additional calls allocate additional stack frames below the current ones. While not mandatory, stacks almost always grow downward, from high addresses to low addresses. Nevertheless, the lowest numerical address occupied on the stack is always called the top of the stack. In addition to their use for holding local variables, stacks can also hold temporary results. The 8088 has an instruction, PUSH, which puts a 16-bit word on top of the stack. This instruction first decrements SP by 2, then stores its operand at the address SP is now pointing to. Similarly, POP removes a 16-bit word from the top of the stack by fetching the value on top of the stack and then incrementing SP by 2. The SP register points to the top of the stack and is modified by PUSH, POP, and CALL instructions, being decremented by PUSH, incremented by POP, and decremented by CALL. The next register in this group is BP, the base pointer. It usually contains an address in the stack. Whereas SP always points to the top of the stack, BP can point to any location within the stack. In practice, a common use for BP is to point to the beginning of the current procedure’s stack frame, in order to make it easy to find the procedure’s local variables. Thus, BP often points to the bottom of the current stack frame (the stack frame word with the highest numerical value) and SP points to the top (the stack frame word with the lowest numerical value). The current stack frame is thus delimited by BP and SP. In this register group, there are two index registers: SI, the source index, and DI, the destination index. These registers are often used in combination with BP to address data in the stack, or with BX to compute the addresses of data memory locations. More extensive treatment of these registers will be deferred to the section on addressing modes. One of the most important registers, which is a group by itself, is the instruction pointer, which is Intel’s name for the program counter (PC). This register is not addressed directly by the instructions, but contains an address in the program code segment of the memory. The processor’s instruction cycle starts by fetching the instruction pointed to by PC. This register is then incremented before the rest of the instruction is executed. In this way this program counter points to the first instruction beyond the current one.

SEC. C.2

THE 8088 PROCESSOR

709

The flag register or condition code register is actually a set of single-bit registers. Some of the bits are set by arithmetic instructions and relate to the result, as follows: Z - result is zero S - result is negative (sign bit) V - result generated an overflow C - result generated a carry A - Auxillary carry (out of bit 3) P - parity of the result

Other bits in this register control operation of certain aspects of the processor. The I bit enables interrupts. The T bit enables tracing mode, which is used for debugging. Finally, the D bit controls the direction of the string operations. Not all 16 bits of this flag register are used; the unused ones are hardwired to zero. There are four registers in the segment register group. Recall that the stack, the data and the instruction codes all reside in main memory, but usually in different parts of it. The segment registers govern these different parts of the memory, which are called segments. These registers are called CS for the code segment register, DS for the data segment register, SS for the stack segment register, and ES for the extra segment register. Most of the time, their values are not changed. In practice, the data segment and stack segment use the same piece of memory, with the data being at the bottom of the segment and the stack being at the top. More about these registers will be explained in Sec. C.3.1.

C.3 MEMORY AND ADDRESSING The 8088 has a somewhat ungainly memory organization due to its combination of a 1-MB memory and 16-bit registers. With a 1-MB memory, it takes 20 bits to represent a memory address. Consequently, it is impossible to store a pointer to memory in any of the 16-bit registers. To get around this problem, memory is organized as segments, each of them 64 KB, so an address within a segment can be represented in 16 bits. We will now go into the 8088 memory architecture in more detail.

C.3.1 Memory Organization and Segments The memory of the 8088, which consists simply of an array of addressable 8bit bytes, is used for the storage of instructions as well as for the storage of data and for the stack. In order to separate the parts of the memory which are used for

710

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

these different purposes, the 8088 uses segments which are chunks of the memory set apart for a certain uses. In the 8088, such a segment consists of 65,536 consecutive bytes. There are four segments: 1. The code segment. 2. The data segment. 3. The stack segment. 4. The extra segment. The code segment contains the program instructions. The contents of the PC register are always interpreted as a memory address in the code segment. A PC value of 0 refers to the lowest address in the code segment, not absolute memory address zero. The data segment contains the initialized and uninitialized data for the program. When BX contains a pointer, it points to this data segment. The stack segment contains local variables and intermediate results pushed on the stack. Addresses in SP and BP are always in this stack segment. The extra segment is a spare segment register that can be placed anywhere in memory that it is needed. For each of the segments, there exists a corresponding segment register: the 16-bit registers CS, DS, SS, and ES. The starting address of a segment is the 20-bit unsigned integer which is constructed by shifting the segment register by 4 bits to the left, and putting zero’s in the four right-most positions. This means that segment registers always indicate multiples of 16, in a 20-bit address space. The segment register points to the base of the segment. Addresses within the segment can be constructed by converting the 16-bit segment register value to its true 20-bit address by appending four zero bits to the end and adding the offset to that. In effect, an absolute memory address is computed by multiplying the segment register by 16 and then adding the offset to it. For example, if DS is equal to 7, and BX is 12, then the address indicated by BX is 7 × 16 + 12 = 124. In other words, the 20-bit binary address implied by DS = 7 is 00000000000001110000. Adding the 16-bit offset 0000000000001100 (decimal 12) to the segment’s origin gives the 20-bit address 00000000000001111100 (decimal 124). For every memory reference, one of the segment registers is used to construct the actual memory address. If some instruction contains a direct address without reference to a register, then this address is automatically in the data segment, and DS is used to determine the base of the segment. The physical address is found by adding this bottom to the address in the instruction. The physical address in memory of the next instruction code is obtained by shifting the contents of CS by four binary places and adding the value of the program counter. In other words, the true 20-bit address implied by the 16-bit CS register is first computed, then the 16-bit PC is added to it to form a 20-bit absolute memory address.

SEC. C.3

MEMORY AND ADDRESSING

711

The stack segment is made up of 2-byte words and so the stack pointer, SP, should always contain an even number. The stack is filled up from high addresses to low addresses. Thus, the PUSH instruction decreases the stack pointer by 2 and then stores the operand in the memory address computed from SS and SP. The POP command retrieves the value, and increments SP by 2. Addresses in the stack segment which are lower than those indicated by SP are considered free. Stack cleanup is thus achieved by merely increasing SP. In practice, DS and SS are always the same, so a 16-bit pointer can be used to refer to a variable in the shared data/stack segment. If DS and SS were different, a 17th bit would be needed on each pointer to distinguish pointers into the data segment from pointers into the stack segment. In retrospect, having a separate stack segment at all was probably a mistake. If addresses in the four segment registers are chosen to be far apart, then the four segments will be disjointed, but if the available memory is restricted, it is not necessary to make them disjoint. After compilation, the size of the program code is known. It is then efficient to start the data and stack segments at the first multiple of 16 after the last instruction. This assumes that the code and data segment will never use the same physical addresses.

C.3.2 Addressing Almost every instruction needs data, either from memory or from the registers. To name this data, the 8088 has a reasonably versatile collection of addressing modes. Many instructions contain two operands, usually called destination and source. Think, for instance, about the copy instruction, or the add instruction: MOV AX,BX

or ADD CX,20

In these instructions, the first operand is destination and the second is the source. (The choice of which goes first is arbitrary; the reverse choice could also have been made.) It goes without saying that, in such a case, the destination must be a left value that is, it must be a place where something can be stored. This means that constants can be sources, but not destinations. In its original design, the 8088 required that at least one operand in a twooperand instruction be a register. This was done so that the difference between word instructions, and byte instructions could be seen by checking whether the addressed register was a word register or a byte register. In the first release of the processor, this idea was so strictly enforced that it was even impossible to push a constant, because neither the source nor the destination was a register in that instruction. Later versions were less strict, but the idea influenced the design

712

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

anyway. In some cases, one of the operands is not mentioned. For example, in the MULB instruction, only the AX register is powerful enough to act as a destination. There are also a number of one-operand instructions, such as increments, shifts, negates, etc. In these cases, there is no register requirement, and the difference between the word and byte operations has to be inferred from the opcodes (i.e., instruction types) only. The 8088 supports four basic data types: 1-byte byte, the 2-byte word, the 4byte long, and binary coded decimal, in which two decimal digits are packed into a word. The latter type is not supported by the interpreter. A memory address always refers to a byte, but in case of a word or a long, the memory locations directly above the indicated byte are implicitly referred to as well. The word at 20 is in the memory locations 20 and 21. The long at address 24 occupies the addresses 24, 25, 26 and 27. The 8088 is little endian, meaning that the low-order part of the word is stored at the lower address. In the stack segment, words should be placed at even addresses. The combination AX DX, in which AX holds the low-order word, is the only provision made for longs in the processor registers. The table of Fig. C-3 gives an overview of the 8088 addressing modes. Let us now briefly discuss them. The topmost horizontal block of the table lists the registers. They can be used as operands in nearly all instructions, both as sources and as destinations. There are eight word registers and eight byte registers. The second horizontal block, data segment addressing, contains addressing modes for the data segment. Addresses of this type always contain a pair of parentheses, to indicate that the contents of the address instead of the value is meant. The easiest addressing mode of this type is direct addressing, in which the data address of the operand is in the instruction itself. Example: ADD CX,(20)

in which the contents of the memory word at address 20 and 21 is added to CX. Memory locations are usually represented by labels instead of by numerical values in the assembly language, and the conversion is made at assembly time. Even in CALL and JMP instructions, the destination can be stored in a memory location addressed by a label. The parentheses around the labels are essential (for the assembler we are using) because ADD CX,20

is also a valid instruction, only it means add the constant 20 to CX, not the contents of memory word 20. In Fig. C-3, the # symbol is used to indicate a numerical constant, label, or constant expression involving a label. In register indirect addressing, the address of the operand is stored in one of the registers BX, SI, or DI. In all three cases the operand is found in the data segment. It is also possible to put a constant in front of the register, in which case the

SEC. C.3

MEMORY AND ADDRESSING

713

################################################################################## ! ! ! ! Mode Operand Examples #! ################################################################################# ! ! ! ! Register addressing ! ! ! ! Byte register ! Byte register ! AH,AL,BH,BL,CH,CL,DH,DL ! ! Word register ! Word register ! AX,BX,CX,DX,SP,BP,SI,DI ! #! ################################################################################# ! ! ! ! Data segment addressing ! ! ! ! Direct address ! Address follows opcode ! (#) ! ! ! ! ! ! Register indirect ! Address in register ! (SI), (DI), (BX) ! ! Register displacement ! Address is register+displ. ! #(SI), #(DI), #(BX) ! ! Register with index ! Address is BX + SI/DI ! (BX)(SI), (BX)(DI) ! ! Register index displacement ! BX + SI DI + displacement ! #(BX)(SI), #(BX)(DI) ! #! ################################################################################# ! ! ! ! Stack segment address ! ! ! ! Base Pointer indirect ! Address in register ! (BP) ! ! ! ! ! ! Base pointer displacement ! Address is BP + displ. ! #(BP) ! ! Address is BP + SI/DI ! (BP)(SI), (BP)(DI) ! ! Base Pointer with index ! BP+SI/DI + displacement ! #(BP)(SI), #(BP)(DI) ! Base pointer index displ. #! ################################################################################# ! ! ! ! ! ! ! ! Immediate data ! Data part of instruction ! # ! Immediate byte/word #! ################################################################################# ! ! ! ! ! ! ! ! Implied address ! Address indirect ( SP ) ! PUSH, POP, PUSHF, POPF ! ! Push/pop instruction ! status flag register ! LAHF, STC, CLC, CMC ! ! Load/store flags ! ! ! ! XLAT BX AL, XLAT Translate ! ! ! ! ! MOVS, CMPS, SCAS ! ! Repeated string instructions ! (SI), (DI), (CX) ! AX, AL ! IN #, OUT # ! ! In / out instructions ! AL, AX, DX ! CBW,CWD ! ! Convert byte, word ! ! ! !################################################################################## Figure C-3. Operand addressing modes. The symbol # indicates a numerical value or label.

address is found by adding the register to the constant. This type of addressing, called register displacement, is convenient for arrays. If, for example, SI contains 5, then the fifth character of the string at the label FORMAT can be loaded in AL by MOVB AL,FORMAT(SI).

The entire string can be scanned by incrementing or decrementing the register in each step. When word operands are used, the register should be changed by two each time. It is also possible to put the base (i.e., lowest numerical address) of the array in the BX register, and keep the SI or DI register for counting. This is called register with index addressing. For example: PUSH (BX)(DI)

fetches the contents of the data segment location whose address is given by the

714

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

sum of the BX and DI registers. This value is then pushed onto the stack. The last two types of addresses can be combined to get register with index and displacement addressing, as in NOT 20(BX)(DI)

which complements the memory word at BX + DI + 20 and BX + DI + 21. All the indirect addressing modes in the data segment also exist for the stack segment, in which case the base pointer BP is used instead of the base register BX. In this way (BP) is the only register indirect stack addressing mode, but more involved modes also exist, up to base pointer indirect with index and displacement −1(BP)(SI) . These modes are valuable for addressing local variables and function parameters, which are stored in stack addresses in subroutines. This arrangement is described further in Sec. C.4.5. All the addresses which comply with the addressing modes discussed up to now can be used as sources and as destinations for operations. Together they are defined to be effective addresses. The addressing mode in the remaining two blocks cannot be used as destinations and are not referred to as effective addresses. They can only be used as sources. The addressing mode in which the operand is a constant byte or word value in the instruction itself is called immediate addressing. Thus, for example, CMP AX,50

compares AX to the constant 50 and sets bits in the flag register, depending on the results. Finally, some of the instructions use implied addressing. For these instructions, the operand or operands are implicit in the instruction itself. For example, the instruction PUSH AX

pushes the contents of AX onto the stack by decrementing SP and then copying AX to the location now pointed to by SP. SP is not named in the instruction itself, however; the mere fact that it is a PUSH instruction implies that SP is used. Similarly, the flag manipulation instructions implicitly use the status flags register without naming it. Several other instructions also have implicit operands. The 8088 has special instructions for moving (MOVS), comparing (CMPS), and scanning (SCAS) strings. With these string instructions, the index registers SI and DI are automatically changed after the operation. This behavior is called auto increment or auto decrement mode. Whether SI and DI are incremented or decremented depends on the direction flag in the status flags register. A direction flag value of 0 increments, whereas a value of 1 decrements. The change is 1 for byte instructions and 2 for word instructions. In a way, the stack pointer is also auto increment and auto decrement: it is decremented by 2 at the start of a PUSH and incremented by 2 at the end of a POP.

SEC. C.4

THE 8088 INSTRUCTION SET

715

C.4 THE 8088 INSTRUCTION SET The heart of every computer is the set of instructions it can carry out. To really understand a computer, it is necessary to have a good understanding of its instruction set. In the following sections, we will discuss the most important of the 8088’s instructions. Some of them are shown in Fig. C-4, where they are divided into 10 groups.

C.4.1 Move, Copy and Arithmetic The first group of instructions is the copy and move instructions. By far, the most common is the instruction MOV, which has an explicit source and an explicit destination. If the source is a register, the destination can be an effective address. In this table a register operand is indicated by an r and an effective address by an e, so this operand combination is denoted by e←r. This is the first entry in the Operands column for MOV. Since, in the instruction syntax, the destination is the first operand and the source is the second operand, the arrow ← is used to indicate the operands. Thus, e←r means that a register is copied to an effective address. For the MOV instruction, the source can also be an effective address and the destination a register, which will be denoted by r←e, the second entry in the Operands column of the instruction. The third possibility is immediate data as source, and effective address as destination, which yields e←#. Immediate data in the table is indicated by the sharp sign (#). Since both the word move MOV and the byte move MOVB exist, the instruction mnemonic ends with a B between parentheses. Thus, the line really represents six different instructions. None of the flags in the condition code register are affected by a move instruction, so the last four columns have the entry ‘‘-’’. Note that the move instructions do not move data. They make copies, meaning that the source is not modified as would happen with a true move. The second instruction in the table is XCHG, which exchanges the contents of a register with the contents of an effective address. For the exchange the table uses the symbol → ←. In this case, there exists a byte version as well as a word version. Thus, the instruction is denoted by XCHG and the Operand field contains r← →e. The next instruction is LEA, which stands for Load Effective Address. It computes the numerical value of the effective address and stores it in a register. Next is PUSH, which pushes its operand onto the stack. The explicit operand can either be a constant (# in the Operands column) or an effective address (e in the Operands column). There is also an implicit operand, SP, which is not mentioned in the instruction syntax. What the instruction does is decrement SP by 2, then store the operand at the location now pointed to by SP. Then comes POP, which removes an operand from the stack to an effective address. The next two instructions, PUSHF and POPF, also have implied operands, the push and pop the flags register, respectively. This is also the case for XLAT

716

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

which loads the byte register AL from the address computed from AL + BX . This instruction allows for rapid lookup in tables of size 256 bytes. Officially defined in the 8088, but not implemented in the interpreter (and thus not listed in Fig. C-4), are the IN and OUT instructions. These are, in fact, move instructions to and from an I/O device. The implied address is always the AX register, and the second operand in the instruction is the port number of the desired device register. In the second block of Fig. C-4 are the addition and subtraction instructions. Each of these has the same three operand combinations as MOV: effective address to register, register to effective address, and constant to effective address. Thus, the Operands column of the table contains r←e, e←r, and e←#. In all four of these instructions, the overflow flag, O, the sign flag, S, the zero flag, Z, and the carry flag, C are all set, based on the result of the instruction. This means, for example, that O is set if the result cannot be correctly expressed in the allowed number of bits, and cleared if it can be. When the largest 16-bit number, 0x7fff (32,767 in decimal), is added to itself, the result cannot be expressed as a 16-bit signed number, so O is set to indicate the error. Similar things happen to the other status flags in these operations. If an instruction has an effect on a status flag, an asterisk (*) is shown in the corresponding column. In the instructions ADC and SBB, the carry flag at the start of the operation is used as an extra 1 (or 0), which is seen as a carry or borrow from the previous operation. This facility is especially useful for representing 32-bit or longer integers in several words. For all additions and subtractions, byte versions also exist. The next block contains the multiplication and division instructions. Signed integer operands require the IMUL and IDIV instructions; unsigned ones use MUL and DIV. The AH : AL register combination is the implied destination in the byte version of these instructions. In the word version, the implied destination is the AX DX register combination. Even if the result of the multiplication is only a word or a byte, the DX or AH register is rewritten during the operation. The multiplication is always possible because the destination contains enough bits. The overflow and carry bits are set when the product cannot be represented in one word, or one byte. The zero and the negative flags are undefined after a multiplication. Division also uses the register combinations DX : AX or AH : AL as the destination. The quotient goes into AX or AL and the remainder into DX or AH. All four flags, carry, overflow, zero and negative, are undefined after a divide operation. If the divisor is 0, or if the quotient does not fit into the register, the operation executes a trap, which stops the program unless a trap handler routine is present. Moreover, it is sensible to handle minus signs in software before and after the divide, because in the 8088 definition the sign of the remainder equals the sign of the dividend, whereas in mathematics, a remainder is always nonnegative. The instructions for binary coded decimals, among which Ascii Adjust for Addition (AAA), and Decimal Ajust for Addition (DAA), are not implemented by the interpreter and not shown in Fig. C-4.

SEC. C.4

THE 8088 INSTRUCTION SET

717

######################################################################################################### ! ! ! ! ! Mnemonic Description Status flags Operands ! ! ! ! O ! S ! Z ! C ! ! ! ! ! !######################################################################################################### ! ! ! - ! ! r ← e, e ← r, e ← # ! ! Move word, byte ! MOV(B) ! ! ! ! ! r→ ! Exchange word ! XCHG(B) ←e - ! ! ! ! ! ! r ← #e ! Load effective address ! LEA - ! ! ! ! ! ! ! ! ! ! ! ! PUSH Push onto stack e, # ! ! ! ! ! ! ! ! e stack from Pop POP ! ! ! ! ! ! ! ! ! ! ! ! Push flags ! PUSHF ! ! ! ! ! ! Pop flags ! POPF - ! ! ! ! ! ! ! ! ! ! ! ! XLAT - ! ! ! ! Translate AL !######################################################################################################### ! ! ! * * * * ! ! r ← e, e ← r, e ← # ! ! Add word ! ADD(B) ! ! ! ! r ← e, e ← r, e ← # ! ! Add word with carry ! ADC(B) * * * ! * ! ! ! ! r ← e, e ← r, e ← # ! ! Subtract word ! SUB(B) * * * * ! ! ! ! ! ! ! ! ! ! ! ! * * * ! SBB(B) * ! r ← e, e ← r, e ← # ! ! Subtract word with borrow !######################################################################################################### ! ! ! ! ! ! e ! Multiply signed ! IMUL(B) * U U * ! ! ! ! ! e ! Multiply unsigned ! MUL(B) * * ! ! U ! U ! ! ! ! ! ! U ! U ! U ! U ! ! e ! Divide signed ! IDIV(B) ! U ! U ! U ! U DIV(B) ! ! ! e ! Divide unsigned !######################################################################################################### ! ! ! ! ! ! Sign extend byte-word ! CBW - ! ! ! ! ! ! ! ! Sign extend word-double ! CWD ! ! ! ! ! ! ! ! ! ! ! * * * ! * ! ! e ! Negate binary ! NEG(B) ! ! ! - ! ! ! e ! Logical complement ! NOT(B) ! ! ! ! ! e ! Increment destination ! INC(B) - ! * * * ! ! ! ! ! e ! Decrement destination ! DEC(B) * * * - ! ! ! ! ! ! ! ! !######################################################################################################### ! ! ! 0 * * 0 ! ! e ← r, r ← e, e ← # ! ! Logical and ! AND(B) ! ! ! ! e ← r, r ← e, e ← # ! ! Logical or ! OR(B) * * 0 ! ! ! 0 ! ! e ← r, r ← e, e ← # ! ! Logical exclusive or ! XOR(B) * * 0 ! ! ! 0 ! ! ! ! ! !######################################################################################################### ! ! ! e * * ! right * * e Logical shift ← 1, ← CL SHR(B) ! ! ! ! ! ! ! ! ! e ← 1, e ← CL ! Arithmetic shift right ! SAR(B) * * ! * * ! ! ! ! ! e ← 1, e ← CL ! shift left ! SAL(B) (=SHL(B)) * * ! * * ! ! ! ! ! e ← 1, e ← CL ! Rotate left ! ROL(B) ! ! ! * * ! ! ! ! ! ! ! ! ! e * right e * Rotate ← 1, ← CL ROR(B) ! ! ! ! ! ! ! ! * ! * ! ! e ← 1, e ← CL ! Rotate left with carry ! RCL(B) ! ! ! ! ! ! e ← 1, e ← CL ! Rotate right with carry !######################################################################################################### RCR(B) * * ! ! ! ! ! ! ! ! ! ! 0 ! e operands e * * ← → r, ← → # 0 TEST(B) Test ! ! ! ! ! ! ! ! * * * ! ← r, e → ←# * ! ! e→ ! Compare operands ! CMP(B) ! ! ! ! ! ! ! Set direction flag (↓) ! STD ! ! ! ! ! ! Clear direction flag (↑) ! CLD - ! ! ! ! ! ! ! ! ! ! ! 1 ! ! ! ! Set carry flag ! STC ! ! ! 0 ! ! ! ! ! Clear carry flag ! CLC ! ! ! CMC * ! ! ! ! Complement carry !######################################################################################################### ! ! ! ! ! Jump back if decremented CX ≥ 0 ! label ! LOOP - ! ! ! ! ! ! ! ! ! ! ! ! - ! ! ! label ! Back if Z=1 and DEC(CX)≥0 ! LOOPZ LOOPE ! ! ! - ! ! ! label ! LOOPNZ LOOPNE ! Back if Z=0 and DEC(CX)≥0 ! ! ! ! ! string instruction !######################################################################################################### - ! REP REPZ REPNZ ! Repeat string instruction ! ! ! ! ! ! ! ! ! ! ! - ! ! ! ! Move word string ! MOVS(B) ! ! ! - ! ! ! ! Load word string ! LODS(B) ! ! ! ! ! ! Store word string ! STOS(B) - ! ! ! ! ! ! ! Scan word string ! SCAS(B) * * * * ! ! ! ! ! ! ! Compare word string ! CMPS(B) ! ! ! * * * ! * ######################################################################################################### ! ! ! ! ! ! ! ! - ! ! ! label ! Jump according conditions ! JCC ! ! ! ! ! e, label ! Jump to label ! JMP - ! ! ! ! ! ! e, label ! Jump to subroutine ! CALL - ! ! ! ! ! ! ! ! ! ! ! ! RET Return from subroutine -, # - ! ! ! ! ! ! ! ! SYS ! ! ! ! System call trap !######################################################################################################### ! ! !

Figure C-4. Some of the most important 8088 instructions.

718

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

C.4.2 Logical, Bit and Shift Operations The next block contains instructions for sign extension, negation, logical complement, increment and decrement. The sign extend operations have no explicit operands, but act on the DX : AX or the AH : AL register combinations. The single operand for the other operations of this group can be found at any effective address. The flags are affected in the expected way in case of the NEG, INC and DEC, except that the carry is not affected in the increment and decrement, which is quite unexpected and which some people regard as a design error. The next block of instructions is the two-operand logical group, all of whose instructions behave as expected. In the shift and rotate group, all operations have an effective address as their destination, but the source is either the byte register CL or the number 1. In the shifts, all four flags are affected; in the rotates, only the carry and the overflow are affected. The carry always gets the bit that is shifted or rotated out of the high-order or low-order bit, depending on the direction of the shift or rotate. In the rotates with carry, RCR, RCL, RCRB, and RCLB, the carry together with the operand at the effective address, constitutes a 17-bit or a 9-bit circular shift register combination, which facilitates multiple word shifts and rotates. The next block of instructions is used to manipulate the flag bits. The main ←) is reason for doing this is to prepare for conditional jumps. The double arrow (→ used to indicate the two operands in compare and test operations, which do not change during the operation. In the TEST operation, the logical AND of the operands is computed to set or clear the zero flag and the sign flag. The computed value itself is not stored anywhere and the operand is unmodified. In the CMP, the difference of the operands is computed and all four flags are set or cleared as a result of the comparison. The direction flag, which determines whether the SI and DI registers should be incremented or decremented in the string instructions, can be set or cleared by STD and CLD, respectively. The 8088 also has a parity flag and an auxiliary carry flag. The parity flag gives the parity of the result (odd or even). The auxiliary flag checks whether overflow was generated in the low (4-bit) nibble of the destination. There are also instructions LAHF and SAHF, which copy the low-order byte of the flag register in AH, and vice versa. The overflow flag is in the high-order byte of the condition code register and is not copied in these instructions. These instructions and flags are mainly used for backward compatibility with the 8080 and 8085 processors.

C.4.3 Loop and Repetitive String Operations The following block contains the instructions for looping. The LOOP instruction decrements the CX register and jumps back to the label indicated if the result is positive. The instructions LOOPZ, LOOPE, LOOPNZ and LOOPNE also test the zero flag to see whether the loop should be aborted before CX is 0.

SEC. C.4

THE 8088 INSTRUCTION SET

719

The destination for all LOOP instructions must be within 128 bytes of the current position of the program counter because the instruction contains an 8-bit signed offset. The number of instructions (as opposed to bytes) that can jumped over cannot be calculated exactly because different instructions have different lengths. Usually, the first byte defines the type of an instruction, and so some instructions take only one byte in the code segment. Often, the second byte is used to define the registers and register modes of the instruction, and if the instructions contain displacements or immediate data, the instruction length can increase to four or six bytes. The average instruction length is typically about 2.5 bytes per instruction, so the LOOP cannot jump further back than about 50 instructions. There also exist some special string instruction looping mechanisms. These are REP, REPZ, and REPNZ. Similarly, the five string instructions in the next block of Fig. C-4 all have implied addresses and all use auto increment or auto decrement mode on the index registers. In all of these instructions, the SI register points into the data segment, but the DI register refers to the extra segment, which is based on ES. Together with the REP instruction, the MOVSB can be used to move complete strings in one instruction. The length of the string is contained in the CX register. Since the MOVSB instruction does not affect the flags, it is not possible to check for an ASCII zero byte during the copy operation by means of the REPNZ, but this can be fixed by using first a REPNZ SCASB to get a sensible value in CX and later a REP MOVSB. This point will be illustrated by the string copy example in Sec. C.8. For all of these instructions, extra attention should be paid to the segment register ES, unless ES and DS have the same value. In the interpreter a small memory model is used, so that ES = DS = SS.

C.4.4 Jump and Call Instructions The last block is about conditional and unconditional jumps, subroutine calls, and returns. The simplest operation here is the JMP instruction. It can have a label as destination or the contents of any effective address. A distinction is made between a near jump and a far jump. In a near jump, the destination is in the current code segment, which does not change during the operation. In a far jump, the CS register is changed during the jump. In the direct version with a label, the new value of the code segment register is supplied in the call after the label, in the effective address version, a long is fetched from memory, such that the low word corresponds to the destination label, and the high word to the new code segment register value. It is, of course, not surprising, that such a distinction exists. To jump to an arbitrary address within a 20-bit address space, some provision has to be made for specifying more than 16 bits. The way it is done is by giving new values for CS and PC.

720

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

Conditional jumps The 8088 has 15 conditional jumps, a few of which have two names (e.g., is the same instruction as JUMP NOT LESS THAN). They are listed in Fig. C-5. All of these allow only jumps with a distance of up to 128 bytes from the instruction. If the destination is not within this range, a jump over jump construction has to be used. In such a construction, the jump with the opposite condition is used to jump over the next instruction. If the next instruction contains an unconditional jump to the intended destination, then the effect of these two instructions is just a longer-ranging jump of the intended type. For example, instead of JUMP GREATER OR EQUAL

JB FARLABEL

we have JNA 1f JMP FARLABEL 1:

In other words, if it is not possible to do JUMP BELOW, then a JUMP NOT ABOVE to a nearby label 1 is placed, followed by an unconditional jump to FARLABEL. The effect is the same, at a slightly higher cost in time and space. The assembler generates these jump over jumps automatically when the destination is expected to be too distant. Doing the calculation correctly is a bit tricky. Suppose that the distance is close to the edge, but some of the intervening instructions are also conditional jumps. The outer one cannot be resolved until the sizes of the inner ones are known, and so on. To be safe, the assembler errs on the side of caution. Sometimes it generates a jump over jump when it is not strictly necessary. It only generates a direct condition jump when it is absolutely certain that the target is within range. Most conditional jumps depend on the status flags, and are preceded by a compare or test instruction. The CMP instructions subtracts the source from the destination operand, sets the condition codes and discards the result. Neither of the operands is changed. If the result is zero or has the sign bit on (i.e., is negative), the corresponding flag bit is set. If the result cannot be expressed in the allowed number of bits, the overflow flag is set. If there is a carry out of the high-order bit, the carry flag is set. The conditional jumps can test all of these bits. If the operands are considered to be signed, the instructions using GREATER THAN and LESS THAN should be used. If they are unsigned, the ones using ABOVE and BELOW should be used.

SEC. C.4

THE 8088 INSTRUCTION SET

721

######################################################### ! Instruction ! Description ! When to jump ! !######################################################### ! ! ! JNA, JBE !######################################################### ! Below or equal ! CF=1 or ZF=1 ! ! JNB, JAE, JNC ! Not below ! CF=0 ! !######################################################### ! ! ! !######################################################### ! Zero, equal ! ZF=1 ! JE, JZ ! ! ! ! JNLE, JG Greater than SF=OF and ZF=0 !######################################################### ! ! ! !######################################################### ! Greater equal ! SF=OF ! JGE, JNL ! ! ! ! JO !######################################################### ! Overflow ! OF=1 ! ! JS ! Sign negative ! SF=1 ! !######################################################### ! ! ! JCXZ !######################################################### ! CX is zero ! CX=0 ! ! JB, JNAE, JC ! Below ! CF=1 ! !######################################################### ! ! ! !######################################################### ! Above ! CF=0&ZF=0 ! JNBE, JA ! ! ! ! JNE, JNZ ! Nonzero, nonequal ! ZF=0 ! !######################################################### ! Less than ! SF≠OF ! ! JL, JNGE #! ######################################################## ! ! ! JLE, JNG ! Less or equal ! SF≠OF or ZF=1 ! !######################################################### ! Nonoverflow ! OF=0 ! ! JNO #! ######################################################## ! ! ! JNS !! Nonnegative !! SF=0 !! !!######################################################### Figure C-5. Conditional jumps.

C.4.5 Subroutine Calls The 8088 has an instruction used to call procedures, usually known in assembly language as subroutines. In the same way as in the jump instructions, there exist near call instructions and far call instructions. In the interpreter, only the near call is implemented. The destination is either a label or can be found at an effective address. Parameters needed in the subroutines have to be pushed onto the stack in reverse order first, as illustrated in Fig. C-6. In assembly language, parameters are usually called arguments, but the terms are interchangeable. Following these pushes the CALL instruction is executed. The instruction starts by pushing the current program counter onto the stack. In this way the return address is saved. The return address is the address at which the execution of the calling routine has to be resumed when the subroutine returns. Next the new program counter is loaded either from the label, or from the effective address. If the call is far, then the CS register is pushed before PC and both the program counter and the code segment register are either loaded from immediate data or from the effective address. This finishes the CALL instruction. The return instruction, RET, just pops the return address from the stack, stores it in the program counter and the program continues at the instruction immediately after the CALL instruction. Sometimes the RET instruction contains a positive number as immediate data. This number is assumed to be the number of bytes of

722

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

! ... BP+8 !################## ! ! BP+6 !################## Argument 2 ! ! BP+4 !################## Argument 1 ! ! Return address ! BP+2 !################## ! ! BP Old BP !################## ! ← BP ! BP−2 ################## Local variable 1 ! ! ! BP−4 !################## Local variable 2 ! BP−6 !################## Local variable 3 ! ! ! BP−8 !################## Temporary result ! ← SP Figure C-6. An example stack.

arguments that were pushed onto the stack before the call; it is added to SP to clean up the stack. In the far variant, RETF, the code segment register is popped after the program counter, as would be expected. Inside the subroutine, the arguments need to be accessible. Therefore the subroutine starts often by pushing the base pointer and copying the current value of SP into BP. This means that the base pointer points to its previous value. Now the return address is at BP + 2 and the first and second arguments can be found at the effective addresses BP + 4 and BP + 6, respectively. If the procedure needs local variables, then the required number of bytes can be subtracted from the stack pointer, and those variables can be addressed from the base pointer with negative offsets. In the example of Fig. C-6, there are three single-word local variables, located at BP − 2, BP − 4, and BP − 6, respectively. In this way, the entire set of current arguments and local variables is reachable through the BP register. The stack is used in the ordinary way to save intermediate results, or for preparing arguments for the next call. Without computing the amount of stack used in the subroutine, the stack can be restored before the return by copying the base pointer into the stack pointer, popping the old BP and finally executing the RET instruction. During a subroutine call, the values of the processor registers sometimes change. It is good practice to use some type of convention such that the calling routine need not be aware of the registers used by the called routine. The simplest way to do this is to use the same conventions for system calls and ordinary subroutines. It is assumed that the AX and DX can change in the called routine. If one of these registers contains valuable information then it is advisable for the calling routine to stack them before pushing the arguments. If the subroutine uses other registers as well, those can be pushed onto the stack immediately at the start of the subroutine, and popped before the RET instruction. In other words, a good convention is for the caller to save AX and DX if they contain anything important, and for the callee to save any other registers it overwrites.

SEC. C.4

THE 8088 INSTRUCTION SET

723

C.4.6 System Calls and System Subroutines In order to separate the tasks of opening, closing, reading, and writing files from assembly programming, programs are run on top of an operating system. To allow the interpreter to run on multiple platforms, a set of seven system calls and five functions are supported by the interpreter. They are listed in Fig. C-7. # #################################################################################### ! Nr ! ! ! Return value ! ! Name Arguments Description #! #################################################################################### ! ! ! ! ! ! 5 ! "OPEN ! *name, 0/1/2 ! file descriptor ! Open file ! ! 8 ! "CREAT ! *name, *mode ! file descriptor ! Create file ! ! 3 ! "READ ! fd, buf, nbytes ! # bytes ! Read nbytes in buffer buf ! ! ! ! ! ! ! ! 4 ! "WRITE ! fd, buf, nbytes ! # bytes ! Write nbytes from buffer buf ! ! 6 ! "CLOSE ! fd ! 0 on success ! close file with fd ! ! 19 ! "LSEEK ! fd, offset(long), 0/1/2 ! position (long) ! Move file pointer ! ! ! ! ! ! ! 1 ! "EXIT #! #################################################################################### ! status ! ! Close files Stop process ! ! 117 ! "GETCHAR ! ! read character ! Read character from std input ! ! 122 ! "PUTCHAR ! char ! write byte ! Write character to std output ! ! ! ! ! ! ! ! 127 ! "PRINTF ! *format, arg ! ! Print formatted on std output ! ! 121 ! "SPRINTF ! buf, *format, arg ! ! Print formatted in buffer buf ! ! 125 ! "SSCANF ! buf, *format, arg ! ! Read arguments from buffer buf ! !# #################################################################################### ! ! ! ! ! Figure C-7. Some UNIX system calls and subroutines available in the interpreter.

These twelve routines can be activated by the standard calling sequence; first push the necessary arguments on the stack in reverse order, then push the call number, and finally execute the system trap instruction SYS without operands. The system routine finds all the necessary information on the stack, including the call number of the required system service. Return values are put either in the AX register, or in the DX : AX register combination (when the return value is a long). It is guaranteed that all other registers will keep their values over the SYS instruction. Also, the arguments will still be on the stack after the call. Since they are not needed any more, the stack pointer should be adjusted after the call (by the caller), unless they are needed for a subsequent call. For convenience, the names of the system calls can be defined as constants at the start of the assembler program, so that they can be called by name instead of by number. In the examples, several system calls will be discussed, so in this section only a minimum of necessary detail is supplied. In these system calls, files are opened either by the OPEN or by the CREAT call. In both cases, the first argument is the address of the start of a string containing the file name. The second argument in the OPEN call is either 0 (if the file should be opened for reading), 1 (if it should be opened for writing), or 2 (for both). If the file should allow writes, and does not exist, it is created by the call. In the CREAT call an empty file is created, with permission set according to the

724

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

second argument. Both the OPEN and the CREAT call return a small integer in the AX register, which is called the file descriptor and which can be used for reading, writing or closing the file. A negative return value means the call failed. At the start of the program, three files are already opened with file descriptors: 0 for standard input, 1 for standard output, and 2 for standard error output. The READ and WRITE calls have three arguments: the file descriptor, a buffer to hold the data, and the number of bytes to transfer. Since the arguments are stacked in reverse order, we first push the number of bytes, then the address of the start of the buffer, then the file descriptor and finally the call number (READ or WRITE ). This order of stacking the arguments was chosen to be the same as the standard C language calling sequence in which read(fd, buffer, bytes);

is implemented by pushing the parameters in the order bytes, buffer, and finally fd. The CLOSE call requires just the file descriptor and returns 0 in AX if the file could be closed successfully. The EXIT call requires the exit status on the stack and does not return. The LSEEK call changes the read/write pointer in an open file. The first argument is the file descriptor. Since the second argument is a long, first the highorder word, then the low word should be pushed onto the stack, even when the offset would fit into a word. The third argument indicates whether the new read/write pointer should be computed relative to the start of the file (case 0), relative to the current position (case 1), or relative to the end of the file (case 2). The return value is the new position of the pointer relative to the start of a file, and can be found as a long in the DX : AX register combination. Now we come to the functions that are not system calls. The GETCHAR function reads one character from standard input, and puts it in AL, with AH being set to zero. On failure, the entire AX word is set to −1. The call PUTCHAR writes a byte on standard output. The return value for a successful write is the byte written; on failure it is −1. The call PRINTF outputs formatted information. The first argument to the call is the address of a format string, which tells how to format the output. The string ‘‘%d’’ indicates that the next argument is an integer on the stack, which is converted to decimal notation when printed. In the same way, ‘‘%x’’ converts to hexadecimal and ‘‘%o’’ converts to octal. Furthermore, ‘‘%s’’ indicates that the next argument is a null-terminated string, which is passed to the call through a memory address on the stack. The number of extra arguments on the stack should match the number of conversion indications in the format string. For example, the call printf(′′x = %d and y = %d\n′′, x, y);

prints the string with the numerical values of x and y substituted for the ‘‘%d’’ strings in the format string. Again, for compatibility with C, the order in which

SEC. C.4

THE 8088 INSTRUCTION SET

725

the arguments are pushed is ‘‘y’’, ‘‘x’’, and finally, the address of the format string. The reason for this convention is that printf has a variable number of parameters, and by pushing them in the reverse order the format string itself is always the last one and thus can be located. If the parameters were pushed from left to right, the format string would be deep in the stack and the printf procedure would not know where to find it. In the call PRINTF , the first argument is the buffer, to receive the output string, instead of standard output. The other arguments are the same as in PRINTF . The SSCANF call is the converse of the PRINTF in the sense that the first argument is a string, which can contain integers in decimal, octal, or hexadecimal notation, and the next argument is the format string, which contains the conversion indications. The other arguments are addresses of memory words to receive the converted information. These system subroutines are very versatile and an extensive treatment of the possibilities is far beyond the scope of this appendix. In Sec. C.8, several examples show how they can be used in different situations.

C.4.7 Final Remarks on the Instruction Set In the official definition of the 8088, there exists a segment override prefix, which facilitates the possibility of using effective addresses from a different segment; that is, the first memory address following the override is computed using the indicated segment register. For example, the instruction ESEG MOV DX,(BX)

first computes the address of BX using the extra segment, and then moves the contents to DX. However, the stack segment, in the case of addresses using SP, and the extra segment, in the case of string instructions with the DI register, cannot be overridden. The segment registers SS, DS and ES can be used in the MOV instruction, but it is impossible to move immediate data into a segment register, and those registers cannot be used in an XCHG operation. Programming with changing segment registers and overrides is quite tricky and should be avoided whenever possible. The interpreter uses fixed segment registers, so these problems do not arise here. Floating-point instructions are available in most computers, sometimes directly in the processor, sometimes in a separate coprocessor, and sometimes only interpreted in the software through a special kind of floating point trap. Discussion of those features is outside the scope of this appendix.

C.5 THE ASSEMBLER We have now finished our discussion of the 8088 architecture. The next topic is the software used to program the 8088 in assembly language, in particular the tools we provide for learning assembly language programming. We will first

726

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

discuss the assembler, then the tracer, and then move on to some practical information for using them.

C.5.1 Introduction Up until now, we have referred to instructions by their mnemonics, that is, by short easy-to-remember symbolic names like ADD and CMP. Registers were also called by symbolic names, such as AX and BP. A program written using symbolic names for instructions and registers is called an assembly language program. To run such a program, it is first necessary to translate it into the binary numbers that the CPU actually understands. The program that converts an assembly language program into binary numbers is the assembler. The output of the assembler is called an object file. Many programs make calls to subroutines that have been previously assembled and stored in libraries. To run these programs, the newlyassembled object file and the library subroutines it uses (also object files) must be combined into a single executable binary file by another program called a linker. Only when the linker has built the executable binary file from one or more object files is the translation fully completed. The operating system can then read the executable binary file into memory and execute it. The first task of the assembler is to build a symbol table, which is used to map the names of symbolic constants and labels directly to the binary numbers that they represent. Constants that are directly defined in the program can be put in the symbol table without any processing. However, labels represent addresses whose values are not immediately obvious. To compute their values, the assembler scans the program line by line in what is called the first pass. During this pass, it keeps track of a location counter usually indicated by the symbol ‘‘.’’, pronounced dot. For every instruction and memory reservation that is found in this pass, the location counter is increased by the size of the memory necessary to contain the scanned item. Thus, if the first two instructions are of size 2 and 3 bytes, respectively, then a label on the third instruction will have numerical value 5. For example, if this code fragment is at the start of a program, the value of l will be 5. MOV AX,6 MOV BX,500 L:

At the start of the second pass, the numerical value of every symbol is known. Since the numerical values of the instruction mnemonics are constants, code generation can now begin. One at a time, instructions are read again and their binary values are written into the object file. When the last instruction has been assembled, the object file is complete.

SEC. C.5

THE ASSEMBLER

727

C.5.2 The ACK-Based Assembler, as88 This section describes the details of the assembler/linker as88, which is provided on the CD-ROM and website and which works with the tracer. This assembler is Amsterdam Compiler Kit (ACK) and is patterned after UNIX assemblers rather than MS-DOS or Windows assemblers. The comment symbol in this assembler is the exclamation mark (!). Anything following an exclamation mark until the end of the line is a comment and does not affect the object file produced. In the same way, empty lines are allowed, but ignored. This assembler uses three different sections, in which the translated code and data will be stored. Those sections are related to the memory segments of the machine. The first is the TEXT section, for the processor instructions. Next is the DATA section for the initialization of the memory in the data segment, which is known at the start of the process. The last is the BSS (Block Started by Symbol), section, for the reservation of memory in the data segment that is not initialized (i.e., initialized to 0). Each of these sections has its own location counter. The purpose of having sections is to allow the assembler to generate some instructions, then some data, then some instructions, then more data, and so on, and then have the linker rearrange the pieces so that all the instructions are together in the text segment and all the data words are together in the data segment. Each line of assembly code produces output for only one section, but code lines and data lines can be interleaved. At run time, the TEXT section is stored in the text segment and the data and BSS sections are stored (consecutively) in the data segment. An instruction or data word in the assembly language program can begin with a label. A label may also appear all by itself on a line, in which case it is as though it appeared on the next instruction or data word. For example, in CMP AX,ABC JE L MOV AX,XYZ L:

L is a label that refers to the instruction of data word following it. Two kinds of labels are allowed. First are the global labels, which are alphanumeric identifiers followed by a colon (:). These must all be unique, and cannot match any keyword or instruction mnemonic. Second, in the TEXT section only, we can have local labels, each of which consists of a single digit followed by a colon (:). A local label may occur multiple times. When a program contains an instruction such as JE 2f

this means JUMP EQUAL forward to the next local label 2. Similarly, JNE 4b

means JUMP NOT EQUAL backward to the closet label 4.

728

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

The assembler allows constants to be given a symbolic name using the syntax identifier = expression

in which the identifier is an alphanumeric string, as in BLOCKSIZE = 1024

Like all identifiers in this assembly language, only the first eight characters are significant, so BLOCKSIZE and BLOCKSIZZ are the same symbol, namely, BLOCKSIZ. Expressions can be constructed from constants, numerical values, and operators. Labels are considered to be constants because at the end of the first pass their numerical values are known. Numerical values can be octal (starting with a 0), decimal, or hexadecimal (starting with 0X or 0x). Hexadecimal numbers use the letters a–f or A–F for the values 10–15. The integer operators are +, −, *, /, and %, for addition, subtraction, multiplication, division and remainder, respectively. The logical operators are &, ˆ, and ~, for bitwise AND, bitwise OR and logical complement (NOT) respectively. Expressions can use the square brackets, [ and ] for grouping. Parentheses are NOT used, to avoid confusion with the addressing modes. Labels in expressions should be handled in a sensible way. Instruction labels cannot be subtracted from data labels. The difference between comparable labels is a numerical value, but neither labels nor their differences are allowed as constants in multiplicative or logical expressions. Expressions which are allowed in constant definitions can also be used as constants in processor instructions. Some assemblers have macro facility, by which multiple instructions can be grouped together and given a name, but as88 does not have this feature. In every assembly language, there are some directives that influence the assembly process itself but which are not translated into binary code. They are called pseudoinstructions. The as88 pseudoinstructions are listed in Fig. C-8. The first block of pseudoinstructions determines the section in which the following lines should be processed by the assembler. Usually such a section requirement is made on a separate line and can be put anywhere in the code. For implementation reasons, the first section to be used must be the TEXT section, then the DATA section, then the BSS section. After these initial references, the sections can be used in any order. Furthermore, the first line of a section should have a global label. There are no other restrictions on the ordering of the sections. The second block of pseudoinstructions contains the data type indications for the data segment. There are four types: .BYTE, .WORD, .LONG, and string. After an optional label and the pseudoinstruction keyword, the first three types expect a comma-separated list of constant expressions on the remainder of the line. For strings there are two keywords, ASCII, and ASCIZ, with the only difference being that the second keyword adds a zero byte to the end of the string. Both require a string between double quotes. Several escapes are allowed in string definitions. These include those of Fig. C-9. In addition to these, any specific character can

SEC. C.5

THE ASSEMBLER

729

################################################################### ! Instruction ! ! Description ################################################################### ! ! ! ! .SECT .TEXT ! Assemble the following lines in the TEXT section ! ################################################################### ! .SECT .DATA ! Assemble the following lines in the DATA section ! ################################################################### ! ! ! ! .SECT .BSS ! Assemble the following lines in the BSS section ! ################################################################### ! ! ! .BYTE Assemble the arguments as a sequence of bytes ################################################################### ! ! ! ! .WORD ! Assemble the arguments as a sequence of words ! ################################################################### ! ! ! ! .LONG ! Assemble the arguments as a sequence of longs ! ################################################################### ! .ASCII "str" ! Store str as ascii an string without a trailing zero byte ! ################################################################### ! ! ! .ASCIZ "str" ! Store str as ascii an string with a trailing zero byte !################################################################### ! ! .SPACE n ! Advance the location counter n positions ! ################################################################### ! ! ! !################################################################### ! Advance the location counter up to an n-byte boundary ! .ALIGN n ! ! ! ! Identifier is an external name ! ! .EXTERN ################################################################### Figure C-8. The as88 pseudoinstructions.

be inserted by a backslash and an octal representation, for example, \377 (at most three digits, no 0 required here). ####################################### ! Escape symbol ! ! Description !####################################### ! ! ! \n ! New line (line feed) ! ####################################### ! \t ! Tab ! ####################################### ! ! ! !####################################### ! Backslash ! \\ ! ! ! \b !####################################### ! Back space ! !####################################### ! Form feed ! \f ! ! ! \r !####################################### ! Carriage return ! ! \" ! Double quote ! ! ####################################### ! ! Figure C-9. Some of the escapes allowed by as88.

The SPACE pseudoinstruction simply requires the location pointer to be incremented by the number of bytes given in the arguments. This keyword is especially useful following a label in the BSS segment to reserve memory for a variable. ALIGN keyword is used to advance the location pointer to the first 2-, 4-, or 8-byte boundary in memory to facilitate the assembly of words, longs, etc. at a suitable memory location. Finally, the keyword EXTERN announces that the routine or memory location mentioned will be made available to the linker for external references. The definition need not be in the current file; it can also be somewhere else, as long as the linker can handle the reference. Although the assembler itself is fairly general, when it is used with the tracer some small points are worth noting. The assembler accepts keywords in either

730

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

uppercase or lowercase but the tracer always displays them in uppercase. Similarly, the assembler accepts both ‘‘\r’’ (carriage return) and ’’\n’’ (line feed) as the new line indication, but the tracer always uses the latter. Moreover, although the assembler can handle programs split over multiple files, for use with the tracer, the entire program must be in a single file with extension ‘‘.$’’. Inside it, include files can be requested by the command #include filename

In this case, the required file is also written in the combined ‘‘.$’’ file at the position of the request. The assembler checks whether the include file was already processed and loads only one copy. This is especially useful if several files use the same header file. In this case, only one copy is included in the combined source file. In order to include the file, the #include must be the first token of the line without leading white space, and the file path must be between double quotes. If there is a single source file, say pr.s, then it is assumed that the project name is pr, and the combined file will be pr.$. If there is more than one source file, then the basename of the first file is taken to be the projectname, and used for the definition of the .$ file, which is generated by the assembler by concatenating the source files. This behavior can be overridden if the command line contains a ‘‘–o projname’’ flag before the first source file, in which case the combined file will be projname. $. Note that there are some drawbacks in using include files and more than one source. It is necessary that the names of labels, variables and constants are different for all sources. Moreover, the file which is eventually assembled to the load file is the projname. $ file, so the line numbers mentioned by the assembler in case of errors and warnings are determined with respect to this file. For very small projects, it is sometimes simplest to put the entire program in one file and avoid #include.

C.5.3 Some Differences with Other 8088 Assemblers The assembler, as88, is patterned after the standard UNIX assembler, and, as such, differs in some ways from the Microsoft Macro Assembler MASM and the Borland 8088 assembler TASM. Those two assemblers were designed for the MS-DOS operating system, and in places the assembler issues and the operating system issues are closely interrelated. Both MASM and TASM support all 8088 memory models allowed by MS-DOS. There is, for example, the tiny memory model, in which all code and data must fit in 64 KB, the small model, in which the code segment and the data segment each can be 64 KB, and large models, which contain multiple code and data segments. The difference between those models depends on the use of the segment registers. The large model allows far calls and changes in the DS register. The processor itself puts some restrictions on

SEC. C.5

THE ASSEMBLER

731

the segment registers, (e.g., the CS register is not allowed as destination in a MOV instruction). To make tracing simpler, the memory model used in as88 resembles the small model, although the assembler without the tracer can handle the segment registers without additional restrictions. These other assemblers do not have a .BSS section, and initialize memory only in the DATA sections. Usually the assembler file starts with some header information, then the DATA section, which is indicated by the keyword .data, followed by the program text after the keyword .code. The header has a keyword title to name the program, a keyword .model to indicate the memory model, and a keyword .stack to reserve memory for the stack segment. If the intended binary is a .com file, then the tiny model is used, all segment registers are equal, and at the head of this combined segment 256 bytes are reserved for a ‘‘Program Segment Prefix.’’ Instead of the .WORD .BYTE and ASCIZ directives, these assemblers have keywords DW for define word and DB for define byte. After the DB directive, a string can be defined inside a pair of double quotes. Labels for data definitions are not followed by a colon. Large chunks of memory are initialized by the DUP keyword, which is preceded by a count and followed by an initialization, for example, the statement LABEL DB 1000 DUP (0)

initializes 1000 bytes of memory with ASCII zero bytes at the label LABEL. Furthermore, labels for subroutines are not followed by a colon, but by the keyword PROC. At the end of the subroutine, the label is repeated and followed by the keyword ENDP, so the assembler can infer the exact scope of a subroutine. Local labels are not supported. The keywords for the instructions are identical in MASM, TASM, and as88. Also, the source is put after the destination in two operand instructions. However, it is common practice to use registers for the passing of arguments to functions, instead of on the stack. If, however, assembly routines are used inside C or C++ programs, then it is advisable to use the stack in order to comply with the C subroutine calling mechanism. This is not a real difference, since it is also possible to use registers instead of the stack for arguments in as88. The biggest difference between the MASM, TASM and as88 is in making system calls. The system is called in MASM and TASM by means of a system interrupt INT. The most common one is INT 21H, which is intended for the MSDOS function calls. The call number is put in AX, so again we have passing of arguments in registers. For different devices there are different interrupt vectors, and interrupt numbers, such as INT 16H for the BIOS keyboard functions and INT 10H for the display. In order to program these functions, the programmer has to be aware of a great deal of device-dependent information. In contrast, the UNIX system calls available in as88 are much easier to use.

732

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

C.6 THE TRACER The tracer-debugger is meant to run on a 24 × 80 ordinary (VT100) terminal, with the ANSI standard commands for terminals. On UNIX or Linux machines, the terminal emulator in the X-window system usually meets the requirements. On Windows machines, the ansi.sys driver usually has to be loaded in the system initialization files as described below. In the tracer examples, we have already seen the layout of the tracer window. As can be seen in Fig. C-10, the tracer screen is subdivided into seven windows. Program t e x t

Processor with

Stack

registers

Source file

Subroutine c a l l stack

E r r o r output field

Interpreter commands

Output field

I n p u t field

V a l u e s of g l o b a l v a r i a b l e s D a t a segment Figure C-10. The tracer’s windows.

The upper left window is the processor window, which displays the general registers in decimal notation and the other registers in hexadecimal. Since the numerical value of the program counter is not very instructive, the position in the program source code with respect to the previous global label is supplied on the line below it. Above the program counter field, five condition codes are shown. Overflow is indicated by a ‘‘v’’, the direction flag by ‘‘>’’ for increasing and by ‘‘’’. Return addresses of subroutines are indicated by a digit in front of the hexadecimal value. The upper right window displays a part of the source file in the neighborhood of the next instruction to be executed. The position of the program counter is also indicated by an arrow ‘‘=>’’.

SEC. C.6

THE TRACER

733

In the window under the processor, the most recent source code subroutine call positions are displayed. Directly under it is the tracer command window, which has the previously-issued command on top and the command cursor on the bottom. Note that every command needs to be followed by a carriage return (labeled Enter on PC keyboards). The bottom window can contain six items of global data memory. Every item starts with a position relative to some label, followed by the absolute position in the data segment. Next comes a colon, then eight bytes in hexadecimal. The next 11 positions are reserved for characters, followed by four decimal word representations. The bytes, the characters, and the words each represent the same memory contents, although for the character representation we have three extra bytes. This is convenient, because it is not clear from the start whether the data will be used as signed or unsigned integers, or as a string. The middle right window is used for input and output. The first line is for error output of the tracer, the second line for input, and then there are some lines left for output. Error output is preceded by the letter ‘‘E’’, input by an ‘‘I’’, and standard output by a ‘‘>’’. In the input line there is an arrow ‘‘->’’ to indicate the pointer which is to be read next. If the program calls read or getchar, the next input in the tracer command line is going into the input field. Also, in this case, it is necessary to close the input line with a return. The part of the line which has not yet processed can be found after the ‘‘->’’ arrow. Usually, the tracer reads both its commands and its input from standard input. However, it is also possible to prepare a file of tracer commands and a file of input lines to be read before the control is passed to the standard input. Tracer command files have extensions .t and input files .i. In the assembly language, both uppercase and lowercase characters can be used for keywords, system subroutines and pseudoinstructions. During the assembly process, a file with extension .$ is made in which those lowercase keywords are translated into uppercase and carriage return characters are discarded. In this way, for each project, say, pr we can have up to six different files: 1. pr.s for the assembly source code. 2. pr.$ for the composite source file. 3. pr.88 for the load file. 4. pr.i for preset standard input. 5. pr.t for preset tracer commands. 6. pr.# for linking the assembly code to the load file. The last file is used by the tracer to fill the upper right window and the program counter field in the display. Also, the tracer checks whether the load file has been created after the last modification of the program source; if not it issues a warning.

734

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

C.6.1 Tracer Commands Figure C-11 lists the tracer commands. The most important ones are the single return command, which is at the first line of the table and which executes exactly one processor instruction, and the quit command q, at the bottom line of the table. If a number is given as a command, then that number of instructions is executed. The number k is equivalent to typing a return k times. The same effect is achieved if the number is followed by an exclamation mark, !, or an X. The command g can be used to go to a certain line in the source file. There are three versions of this command. If it is preceded by a line number, then the tracer executes until that line is encountered. With a label /T, with or without +#, the line number at which to stop is computed from the instruction label T. The g command, without any indication preceding it, causes the tracer to execute commands until the current line number is again encountered. ########################################################################## ! ! Address ! Command ! Example ! Description #! ######################################################################### ! ! ! ! ! ! Execute one instruction ! ! #! ######################################################################### ! ! Execute # instructions ! 24 ! ,!,X ! # #! ######################################################################### ! ! ! ! ! ! /start+5g ! Run until line # after label T ! g,!, /T+# #! ######################################################################### ! ! ! ! ! T label # after line on breakpoint Put /start+5b /T+# b #! ######################################################################### ! ! ! ! ! /start+5c ! Remove breakpoint on line # after label T ! ! c !########################################################################## /T+# ! ! ! ! ! # ! ! Execute program until line # ! 108g ! g #! ######################################################################### ! Execute program until current line again ! ! g ! g ! #! ######################################################################### ! ! ! ! ! ! Put breakpoint on current line ! b ! b #! ######################################################################### ! ! Remove breakpoint on current line ! c ! c ! #! ######################################################################### ! ! ! ! ! ! Execute program until next line ! n ! n #! ######################################################################### ! ! ! ! ! #! ######################################################################### ! ! Execute until breakpoint or end ! r ! r ! Run program until same subroutine level ! ! = ! = ! #! ######################################################################### ! ! ! ! ! ! Run until subroutine level minus 1 ! ! #! ######################################################################### ! ! Run until subroutine level plus 1 ! + ! + ! #! ######################################################################### ! ! ! ! /D+# ! ! Display data segment on label+# ! /buf+6 ! #! ######################################################################### ! ! ! ! ! /D+# #! ######################################################################### ! ! /buf+6d ! Display data segment on label+# ! d,! ! ! Refresh windows ! R , CTRL L ! R #! ######################################################################### ! ! ! ! ! ! ! Stop tracing, back to command shell ! q ! q #! ######################################################################### Figure C-11. The tracer commands. Each command must be followed by a carriage return (the Enter key). An empty box indicates that just a carriage return is needed. Commands with no Address field listed above have no address. The # symbol represents an integer offset.

The command /label is different for an instruction label and a data label. For a data label, a line in the bottom window is filled or replaced with a set of data

SEC. C.6

THE TRACER

735

starting with that label. For an instruction label, it is equivalent to the g command. The label may be followed by a plus sign and a number (indicated by # in Fig. C-11), to obtain an offset from the label. It is possible to set a breakpoint at an instruction. This is done with the command b, which can be optionally preceded by an instruction label, possibly with an offset. If a line with a breakpoint is encountered during execution, the tracer stops. To start again from a breakpoint, a return or run command is required. If the label and the number are omitted, then the breakpoint is set at the current line. The breakpoint can be cleared by a breakpoint clear command, c, which can be preceded by labels and numbers, like the command b. There is a run command, r, in which the tracer executes until either a breakpoint, an exit call, or the end of the commands is encountered. The tracer also keeps track of the subroutine level at which the program is running. This is shown in the window below the processor window and can also be seen through the indication numbers in the stack window. There are three commands that are based on these levels. The − command causes the tracer to run until the subroutine level is one less than the current level. What this command does is execute instructions until the current subroutine is finished. The converse is the + command, which runs the tracer until the next subroutine level is encountered. The = command runs until the same level is encountered, and can be used to execute a subroutine at the CALL command. If = is used, the details of the subroutine are not shown in the tracer window. There is a related command, n, which runs until the next line in the program is encountered. This command is especially useful when issued as a LOOP command; execution stops exactly when the bottom of the loop is executed.

C.7 GETTING STARTED In this section, we will explain how to use the tools. First of all, it is necessary to locate the software for your platform. We have precompiled versions for Solaris, UNIX, for Linux and for Windows. The tools are located on the CDROM and on the Web at www.prenhall.com/tanenbaum. Once there, click on the Companion Web Site for this book and then click on the link in the left-hand menu. Unpack the selected zip file to a directory assembler. This directory and its subdirectories contain all the necessary material. On the CD-ROM, the main directories are Bigendnx, LtlendNx, and MSWindos, and in each there is a subdirectory assembler which contains the material. The three top-level directories are for Big-Endian UNIX (e.g. Sun workstations), Little-Endian UNIX (e.g., Linux on PCs), and Windows systems, respectively. After unpacking or copying, the assembler directory should contain the following subdirectories and files: READ"ME, bin, as"src, trce"src, examples, and exercise. The precompiled sources can be found in the bin directory but, for convenience, there is also a copy of the binaries in the examples directory.

736

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

To get a quick preview of how the system works, go to the examples directory and type the command t88 HlloWrld

This command corresponds to the first example in Sec. C.8. The source code for the assembler is in the directory as"src. The source code files are in the language C, and the command make should recompile the sources. For POSIX-compliant platforms, there is a Makefile in the source directory which does the job. For Windows, there is a batch file make.bat. It may be necessary to move the executable files after compilation to a program directory, or to change the PATH variable to make the assembler as88 and the tracer t88 visible from the directories containing the assembly source codes. Alternatively, instead of typing t88, the full path name can be used. On Windows 2000 and XP systems, it is necessary to install the ansi.sys terminal driver by adding the line device=%systemRoot%\System32\ansi.sys

to the configuration file, config.nt. The location of this file is as follows: Windows 2000: \winnt\system32\config.nt Windows XP: \windows\system32\config.nt

On Windows 95, 98, and ME, the driver should be added to config.sys. On UNIX and Linux systems, the driver is usually standard.

C.8 EXAMPLES In Sec. C.2 through Sec. C.4, we discussed the 8088 processor, its memory, and its instructions. Then, in Sec. C.5, we studied the as88 assembly language used in this tutorial. In Sec. C.6 we studied the tracer. Finally, in Sec. C.7, we described how to set up the toolkit. In theory, this information is sufficient to write and debug assembly programs with the tools provided. Nevertheless, it may be helpful for many readers to see some detailed examples of assembly programs and how they can be debugged with the tracer. That is the purpose of this section. All the example programs discussed in this section are available in the examples directory in the toolkit. The reader is encouraged to assemble and trace each one as it is discussed.

C.8.1 Hello World Example Let us start with the example of Fig. C-12, HlloWrld.s. The program is listed in the left window. Since the assembler’s comment symbol is the exclamation mark (!), it is used in the program window to separate the instructions from the

SEC. C.8

737

EXAMPLES

line numbers that follow. The first three lines contain constant definitions, which connect the conventional names of two system calls and the output file to their corresponding internal representations. The pseudoinstruction .SECT, on line 4, states that the following lines should be considered to be part of the TEXT section; that is, processor instructions. Similarly, line 17 indicates that what follows is to be considered data. Line 19 initializes a string of data consisting of 12 bytes, including one space and a line feed (\n) at the end. Lines 5, 18 and 20 contain labels, which are indicated by a colon :. These labels represent numerical values, similar to constants. In this case, however, the assembler has to determine the numerical values. Since start is at the beginning of the TEXT section, its value will be 0, but the value of any subsequent labels in the TEXT section (not present in this example), would depend on how many bytes of code preceded them. Now consider line 6. This line ends with the difference of two labels, which is numerically a constant. Thus, line 6 is effectively the same as MOV CX,12

except that it lets the assembler determine the string length, rather than making the programmer do it. The value indicated here is the amount of space in the data reserved for the string on line 19. The MOV on line 6 is the copy command, which requires the de −hw to be copied to CX. _EXI T = 1 _WRITE = 4 _STDOUT = 1 .SECT .TEXT start: MOV CX,de-hw PUSH CX PUSH hw PUSH _STDOUT PUSH _WRITE SYS ADD SP, 8 SUB CX,AX PUSH CX PUSH _EXIT SYS .SECT .DATA hw: . ASCII "Hello World\n " de: .BYTE 0

! 1 ! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13 ! 14 ! 15 ! 16 ! 17 ! 18 ! 19 ! 20

C S : 00 D S = S S = E S : 002 AH:00 A L : 0 c AX: 12 BH:00 BL:00 B X : 0 CH:00 C L : 0 c C X : 12 DH:00 D L : 0 0 DX: 0 S P : 7 f d 8 SF O D S Z C => 0004 B P : 0000 CC - > p - 0001 S I : 0000 IP:000c:PC 0000 D I : 0000 start + 7 000c E I

=>

MOV PUSH PUSH PUSH PUSH SYS ADD SUB PUSH

CX,de-hw CX HW _STDOUT _WRITE SP , 8 CX,AX CX

! ! ! ! ! ! ! ! !

6 7 8 9 10 11 12 13 14

hw

> H e l l o World\n hw + 0 = 0 0 0 0 : 48 65 6c 6c 6f 20 57 6f H e l l o

World

25928

(b)

(a)

Figure C-12. (a) HlloWrld.s. (b) The corresponding tracer window.

Lines 7 through 11 show how system calls are made in the toolkit. These five lines are the assembly code translation of the C language function call write(1, hw, 12);

738

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

where the first parameter is the file descriptor for standard output (1), the second is the address of the string to be printed (hw), and the third is the length of the string (12). Lines 7 through 9 push these parameters onto the stack in reverse order, which is the C calling sequence and the one used by the tracer. Line 10 pushes the system call number for write (4) onto the stack, and line 11 makes the actual call. While this calling sequence closely mimics how an actual assembly language program would work on a UNIX (or Linux) PC, for a different operating system, it would have to be modified slightly to use the calling conventions of that operating system. The as88 assembler and t88 tracer use the UNIX calling conventions even when they are running on Windows, however. The system call on line 11 does the actual printing. Line 12 performs a cleanup on the stack, resetting the stack pointer back to the value it had before the four 2-byte words were pushed onto the stack. If the write call is successful, the number of bytes written is returned in AX. Line 13 subtracts the system call result after line 11 from the original string length in CX to see whether the call was successful, that is, to see if all the bytes were written. Thus, the exit status of the program will be 0 on success and something else on failure. Lines 14 and 15 prepare for the exit system call on line 16 by pushing the exit status and function code for the EXIT call onto the stack. Note that in the MOV and SUB instructions the first argument is the destination and the second is the source. This is the convention used by our assembler; other assemblers may reverse the order. There is no particular reason to choose one order over the other. Now let us try to assemble and run HlloWrld.s. Instructions will be given for both UNIX and Windows platforms. For Linux, Solaris, MacOS X, and other UNIX variants, the procedure should be essentially the same as for UNIX. First, start up a command prompt (shell) window. On Windows, the click sequence is usually Start > Programs > Accessories > Command prompt

Next, change to the examples directory using the cd (Change directory) command. The argument to this command depends on where the toolkit has been placed in the file system. Then verify that the assembler and tracer binaries are in this directory, using ls on UNIX and dir on Windows systems. They are called as88 and t88, respectively. On Windows systems, they have the extension .exe, but that need not be typed in the commands. If the assembler and tracer are not there, find them and copy them there. Now assemble the test program using as88 HlloWrld.s

If the assembler is present in the examples directory but this command gives an error message, try typing ./as88 HlloWrld.s

SEC. C.8

EXAMPLES

739

on UNIX systems or .\as88 HlloWrld.s

on Windows systems. If the assembly process completes correctly, the following messages will be displayed: Project HlloWrld listfile HlloWrld.$ Project HlloWrld num file HlloWrld.# Project HlloWrld loadfile HlloWrld.88.

and the corresponding three files created. If there are no error messages, give the tracer command: t88 HlloWrld

The tracer display will appear with the arrow in the upper right-hand window pointing to the MOV CX,de-hw

instruction of line 6. Now hit the return (called Enter on PC keyboards) key. Notice that the instruction pointed to is now PUSH CX

and the value of CX in the left-hand window is now 12. Hit return again and notice that the middle window on the top line now contains the value 000c, which is hexadecimal for 12. This window shows the stack, which now has one word containing 12. Now hit return three more times to see the PUSH instructions on lines 8, 9, and 10 being carried out. At this point, the stack will have four items and the program counter in the left-hand window will have the value 000b. The next time return is hit, the system call is executed and the string ‘‘Hello World\n’’ is displayed in the lower right-hand window. Note that SP now has the value 0x7ff0. After the next return, SP is incremented by 8 and becomes 0x7ff8. After four more returns, the exit system call completes and the tracer exits. To be certain that you understand how everything works, fetch the file hlloWrld.s into your favorite editor. It is better not to use a word processor. On UNIX systems, ex, vi, or emacs are good choices. On Windows systems, notepad is a simple editor, usually reachable from Start > Programs > Accessories > Notepad

Do not use Word since the display will not look right and the output may be formatted incorrectly. Modify the string on line 19 to display a different message, then save the file, assemble it, and run it with the tracer. You are now starting to do assembly language programming.

740

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

C.8.2 General Registers Example The next example demonstrates in more detail how the registers are displayed and one of the pitfalls of multiplication on the 8088. In Fig. C-13, part of the program genReg.s is shown on the left. To its right are two tracer register windows, corresponding to different stages of the program’s execution. Fig. C-13(b) shows the register state after line 7 has been executed. The instruction MOV AX,258

on line 4 loads the value 258 in AX, which results in the value 1 being loaded into and the value 2 being loaded into AL. Then line 5 adds AL to AH, making AH equal to 3. On line 6, the contents of the variable times (10) are copied into CX. On line 7, the address of the variable muldat, which is 2 because it is at the second byte of the DATA segment, is loaded into BX. This is the instant in time at which the dump of Fig. C-13(b) was made. Note that AH is 3, AL is 2, and AX is 770, which is to be expected, as 3 × 256 + 2 = 770.

AH

start: MOV AX,258 ADDB AH,AL MOV CX,(times) MOV BX,muldat MOV AX,(BX) llp: MUL 2(BX) LOOP llp .SECT .DATA times: .WORD 10 muldat :.WORD 625, 2 (a)

! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13

C S : 00 D S = S S = E S :0 02 AH:03 A L : 0 2 A X : 770 BH:00 BL:02 B X : 2 CH:00 C L : 0 a C X : 10 DH: 0 0 D L : 0 0 D X : 0 S P : 7 f e 0 SF O D S Z C B P : 0000 CC - > p - S I : 0000 I P : 0 0 0 9 : P C D I : 0000 s t a r t + 4 (b)

C S : 00 D S = S S = E S :0 02 A H : 3 8 A L : 8 0 A X : 14464 BH:00 BL:02 B X : 2 CH:00 C L : 0 4 C X : 4 D H :00 D L : 0 1 D X : 1 S P : 7 f e 0 SF O D S Z C B P : 0000 CC v > p - c S I : 0000 I P : 0 0 1 1 : P C D I : 0000 s t a r t + 7 (c)

Figure C-13. (a) Part of a program. (b) The tracer register window after line 7 has been executed. (c)

The next instruction (line 8) copies the contents of muldat into AX. Thus, after the return key is hit, AX will be 625. We are now ready to enter a loop that multiplies the contents of AX by the word addressed by 2BX (i.e., muldat + 2), which has the value 2. The implied destination of the MUL instruction is the DX : AX long register combination. In the first iteration of the loop, the result fits in one word, so AX contains the result (1250), and DX remains 0. The contents of all the registers after 7 multiplications are shown in Fig. C-13. Since AX started at 625, the result after those seven multiplications by 2 is 80,000. This result does not fit in AX, but the product is held in the 32-bit register formed by the concatenation of DX : AX, so DX is 1 and AX is 14,464. Numerically, this value is 1 × 65,536 + 14,464, which is, indeed, 80,000. Note that CX is 4 here, because the LOOP instruction decrements it every iteration. Because it started at 10, after seven executions of the MUL instruction (but only six iterations of the LOOP instruction) we have CX set to 4.

SEC. C.8

741

EXAMPLES

"EXIT =1 "PRINTF = 127 .SECT .TEXT inpstart: MOV BP,SP PUSH vec2 PUSH vec1 MOV CX,vec2-vec1 SHR CX,1 PUSH CX CALL vecmul MOV (inprod),AX PUSH AX PUSH pfmt PUSH "PRINTF SYS ADD SP,12 PUSH 0 PUSH "EXIT SYS

! 1 ! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13 ! 14 ! 15 ! 16 ! 17 ! 18 ! 19 ! 20

define the value of "EXIT define the value of "PRINTF start the TEXT segment define label inpstart save SP in BP push address of vec2 push address of vec1 CX = number of bytes in vector CX = number of words in vector push word count call vecmul move AX push result to be printed psuh address of format string push function code for PRINTF call the PRINTF function clean up the stack push status code push function code for EXIT call the EXIT function

vecmul: PUSH BP MOV BP,SP MOV CX,4(BP) MOV SI,6(BP) MOV DI,8(BP) PUSH 0

! 21 ! 22 ! 23 ! 24 ! 25 ! 26 ! 27

start of vecmul(count, vec1, vec2) save BP on stack copy SP into BP to access arguments put count in CX to control loop SI = vec1 DI = vec2 push 0 onto stack

1:

! 28 ! 29 ! 30 ! 31 ! 32 ! 33 ! 34 ! 35

move (SI) to AX multiply AX by (DI) Add AX to accumulated value in memory Increment DI to point to next element if CX > 0, go back to label 1b Pop top of stack to AX Restore BP Return from subroutine

LODS MUL (DI) ADD -2(BP),AX ADD DI,2 LOOP 1b POP AX POP BP RET

.SECT .DATA pfmt: .ASCIZ "Inner product is: %d\n" .ALIGN 2 vec1:.WORD 3,4,7,11,3 vec2:.WORD 2,6,3,1,0 .SECT .BSS inprod: .SPACE 2

! 36 ! 37 ! 38 ! 39 ! 40 ! 41 ! 42

start DATA segment define string force address even vector 1 vector 2 start BSS segment allocate space for inprod

Figure C-14. The program vecprod.s.

In the next multiplication, trouble crops up. Multiplication involves AX but not DX, so the MUL multiples AX (14464) by 2 to get 28,928. This results in AX being set to 28,928 and DX being set to 0, which is numerically incorrect.

742

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

C.8.3 Call Command and Pointer Registers The next example, vecprod.s is a small program that computes the inner product of two vectors, vec1 and vec2. It is listed in Fig. C-14. The first part of the program prepares to call vecmul by saving SP in BP and then pushing the addresses of vec2 and vec1 onto the stack so that vecmul will have access to them. Then the length of the vector in bytes is loaded in CX on line 8. By shifting this result right one bit, on line 9, CX now contains the number of words in the vector, which is pushed onto the stack on line 10. The call to vecmul is made on line 11. Once again, it is worth mentioning that the arguments of subroutines are, by convention, pushed onto the stack in reverse order to be compatible with the C calling convention. In this way, vecmul can also be called from C using vecmul(count, vec1, vec2)

During the CALL instruction, the return address is pushed onto the stack. If the program is traced, then this address turns out to be 0x0011. The first instruction in the subroutine is a PUSH of the base pointer, BP, on line 22. BP is saved because we will need this register to address the arguments and the local variables of the subroutine. Next, the stack pointer is copied to the BP register on line 23, so that the new value of the base pointer is pointing to the old value. Now everything is ready for loading the arguments into registers and for reserving space for a local variable. In the next three lines, each of the arguments is fetched from the stack and put in a register. Recall that the stack is word oriented, so stack addresses should be even. The return address is next to the old base pointer so it is addressed by 2(BP). The count argument is next and addressed by 4(BP). It is loaded into CX on line 24. In lines 25 and 26, SI is loaded with vec1 and DI is loaded with vec2. This subroutine needs one local variable with initial value 0 to save the intermediate result. To this end, the value 0 is pushed on line 27. The state of the processor just before the loop is entered on line 28 for the first time is shown in Fig. C-15. The narrow window in the middle of the top row (to the right of the registers) shows the stack. At the bottom of the stack is the address of vec2 (0x0022), with vec1 (0x0018) above it and the third argument, the number of items in each vector (0x0005) above that. Next comes the return address (0x0011). The number 1 to the left of this address indicates it is a return address one level from the main program. In the window below the registers, the same number 1 is shown, this time giving its symbolic address. Above the return address in the stack is the old value of BP (0x7fc0) and then the zero pushed on line 27. The arrow pointing to this value indicates where SP points. The window to the right of the stack shows a fragment of the program text, with the arrow indicating the next instruction to be executed.

SEC. C.8

743

EXAMPLES

MOV BP,SP PUSH vec2 PUSH vec1 MOV CX,vec2-vec1 SHR CX,1 PUSH CX CALL vecmul ---------vecmul : PUSH BP MOV BP,SP MOV CX,4(BP) MOV SI,6(BP) MOV DI,8(BP) PUSH 0 1: LODS MUL (DI) ADD -2(BP),A X ADD DI,2 LOOP 1b

! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 21 ! 22 ! 23 ! 24 ! 25 ! 26 ! 27 ! 28 ! 29 ! 30 ! 31 ! 32

C S : 00 D S = S S = E S :0 04 AH:00 A L : 0 0 A X : 0 BH:00 BL:00 B X : 0 CH:00 C L : 0 5 C X : 5 =>0000 DH:00 D L : 0 0 DX: 0 7fc0 S P : 7 f b 4 SF O D S Z C 1 0011 B P : 7 f b 6 CC - > p z 0005 = > 1 : S I : 0018 I P : 0 0 3 1 : P C 0018 D I : 0022 vecmul+7 0022 1 =0018: 3 0 4 0 7 0 b 0 . . . . . . . . . . . 3 =0022: 2 0 6 0 3 0 1 0 . . . . . . . . . . . 2 = 0 0 0 0 :54 68 65 20 69 6e 20 70 The in p r o d 26708 = 0 0 1 2 :25 64 21 a 0 0 3 0 % d ! . . . . . . . . 25637

Figure C-15. Execution of vecprod.s when it reaches line 28 for the first time.

Now let us examine the loop starting at line 28. The instruction LODS loads a memory word indirectly through the register SI from the data segment into AX. Because the direction flag is set, LODS is in auto-increment mode, so after the instruction SI will point to the next entry of vec1. To see this effect graphically, start the tracer with the command t88

vecprod

When the tracer window appears, type the command /vecmul+7b

followed by a return to put a breakpoint at the line containing the LODS. From now on, we will not mention that all commands must be followed by the return key. Then give the command g

to have the tracer execute commands until the breakpoint is encountered. It will stop at the line containing the LODS. On line 29, the value of AX is multiplied to the source operand. The memory word for the MUL instruction is fetched from the data segment through the DI in register indirect mode. The implied destination of MUL is the DX : AX long register combination which is not mentioned in the instruction but which is implied by it. On line 30, the result is added to the local variable at the stack address −2(BP). Because MUL does not autoincrement its operand, that must be done explicitly on line 31. Afterward, DI points to the next entry of vec2. The LOOP instruction finishes this step. Register CX is decremented, and, if it is still positive, the program jumps back to the local label 1 on line 28. The use of

744

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

the local label 1b means the closest label 1 looking backward from the current location. After the loop, the subroutine pops the return value into AX (line 33), restores BP (line 34), and returns to the calling program (line 35). Then the main program is resumed after the call with the MOV instruction on line 12. This instruction is the start of a five-instruction sequence whose goal is to print the result. The printf system call is modeled after the printf function in the standard C programming library. Three arguments are pushed onto the stack on lines 13-15. These arguments are the integer value to be printed, the address of the format string (pfmt), and the function code for printf (127). Note that the format string pfmt contains a %d to indicate that an integer variable can be found as argument to the printf call to complete the output. Line 17 cleans up the stack. Since the program started on line 5 by saving the stack pointer in the base pointer, we could also use the instruction MOV SP,BP

for a stack cleanup. The advantage of this solution is that the programmer does not need to keep the stack balanced in the process. For the main program this is not a big issue, but in subroutines this approach is an easy way to throw away garbage such as obsolete local variables. The subroutine vecmul can be included in other programs. If the source file vecprod.s is put on the command line behind another assembler source file, the subroutine is available for multiplication two vectors of a fixed length. It is advisable to remove the constant definitions "EXIT and "PRINTF first, in order to avoid their being defined twice. If the header file syscalnr.h is included somewhere, then there is no need to define the system call constants anywhere else.

C.8.4 Debugging an Array Print Program In the previous examples, the programs examined were simple but correct. Now we will show how the tracer can help debug incorrect programs. The next program is supposed to print the integer array, which is supplied after the label vec1. However, the initial version contains three errors. The assembler and tracer will be used to correct those errors, but first we will discuss the code. Because every program needs system calls, and thus must define constants by which to identify the call numbers, we have put the constant definitions for those numbers in a separate header file ../syscalnr.h, which is included on line 1 of the code. This file also defines the constants for the file descriptors STDIN = 0 STDOUT = 1 STDERR = 2

which are opened at the start of the process, and header labels for the text and the data segments. It is sensible to include it at the head of all assembly source files,

SEC. C.8

745

EXAMPLES

as these are much used definitions. If a source is distributed over more than one file, the assembler includes only the first copy of this header file, to avoid defining the constants more than once. The program arrayprt is shown in Fig. C-16. The comments have been omitted here, as the instructions should be well known by now. In this way, a twocolumn format can be used. Line 4 puts the address of the empty stack in the base pointer register to allow the stack cleanup can be made on line 10 by copying the base pointer to the stack pointer, as described in the previous example. We also have seen the computation and pushing of the stack arguments before the call on lines 5 through 9 in the previous example. Lines 22 to 25 load the registers in the subroutine. #include "../syscalnr.h"

! 1

.SECT .TEXT vecpstrt: MOV BP,SP PUSH vec1 MOV CX,frmatstr-vec1 SHR CX PUSH CX CALL vecprint MOV SP,BP PUSH 0 PUSH "EXIT SYS

! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13

.SECT .DATA vec1: .WORD 3,4,7,11,3 frmatstr: .ASCIZ "%s"

! 14 ! 15 ! 16

frmatkop: ! 17 .ASCIZ "The array contains " ! 18 frmatint: .ASCIZ " %d" ! 19

.SECT .TEXT vecprint: PUSH BP MOV BP,SP MOV CX,4(BP) MOV BX,6(BP) MOV SI,0 PUSH frmatkop PUSH frmatstr PUSH "PRINTF SYS MOV -4(BP),frmatint 1: MOV DI,(BX)(SI) MOV -2(BP),DI SYS INC SI LOOP 1b PUSH ’\n’ PUSH "PUTCHAR SYS MOV SP,BP RET

! 20 ! 21 ! 22 ! 23 ! 24 ! 25 ! 26 ! 27 ! 28 ! 29 ! 30 ! 31 ! 32 ! 33 ! 34 ! 35 ! 36 ! 37 ! 38 ! 39 ! 40 ! 41

Figure C-16. The program arrayprt before debugging.

Lines 27 to 30 show how a string can be printed, and 31 to 34 show the printf system call for an integer value. Note that the address of the string is pushed on line 27, while on line 33 the value of the integer is moved onto the stack. In both cases the address of the format string is the first argument of PRINTF. Lines 37 to 39 show how a single character can be printedusing the putchar system call. Now let us try assembling and running the program. When the command as88 arrayprt.s

is typed, we get an operand error on line 28 of the file arrayprt.$ This file is generated by the assembler by combining the included files with the source file to get

746

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

a composite file that is the actual assembler input. To see where line 28 really is, we have to examine line 28 of arrayprt.$. We cannot look at arrayprt.s to get the line number because the two files do not match on account of the header being included line by line in arrayprt.$ Line 28 in arrayprt.$ corresponds to line 7 in arrayprt.s because the included header file, syscalnr.h, contains 21 lines. One easy way to find line 28 of arrayprt.$ on UNIX systems is to type the command head –28 arrayprt.$

which displays the first 28 lines of the combined file. The line at the bottom of the listing is the one in error. In this way (or by using an editor and going to line 28) we see that the error is on line 7, which contains the SHR instruction. Comparing this code with the instruction table in Fig. C-4 shows the problem: the shift count has been omitted. The corrected line 7 should read SHR CX,1

It is very important to note that the error must be corrected in the original source file, arrayprt.s, and not in the combined source, arrayprt.$, as the latter is automatically regenerated every time the assembler is called. The next attempt to assemble the source code file should succeed. Then the tracer can be started by the command: t88 arrayprt

During the tracing process, we can see that the output is not consistent with the vector in the data segment. The vector contains: 3, 4, 7, 11, 3, but the values displayed start with: 3, 1024, ... . Clearly, something is wrong. To find the error, the tracer can be run again, step by step, examining the state of the machine just before the incorrect value is printed. The value to be printed is stored in memory on lines 32 and 33. Since the wrong value is being printed, this is a good place to see what is wrong. The second time through the loop, we see that SI is an odd number, when clearly it should be an even number, as it is indexing through words, not bytes. The problem is on line 35. It increments SI by 1; it should increment it by 2. To fix the bug, this line should be changed to ADD SI,2

When this correction is made, the printed list of numbers is correct. However, there is one more error waiting for us. When vecprint is finished and returns, the tracer complains about the stack pointer. The obvious thing to check for now is whether the value pushed onto the stack when vecprint is called is the value on top of the stack when the RET on line 41 is executed. It is not. The solution is to replace line 40 with two lines: ADD SP,10 POP BP

SEC. C.8

747

EXAMPLES

the first instruction removes the 5 words pushed onto the stack during vecprint, thus exposing the value of BP saved on line 22. By popping this value to BP, we restore BP to its precall value and expose the correct return address. Now the program terminates correctly. Debugging assembly code is definitely more of an art than a science, but the tracer makes it much easier than running on the bare metal. .SECT .TEXT stcstart: PUSH mesg1 PUSH mesg2 CALL strngcpy ADD SP,4 PUSH 0 PUSH 1 SYS strngcpy: PUSH CX PUSH SI PUSH DI PUSH BP MOV BP,SP MOV AX,0 MOV DI,10(BP) MOV CX,−1 REPNZ SCASB NEG CX DEC CX MOV SI,10(BP) MOV DI,12(BP) PUSH DI REP MOVSB CALL stringpr MOV SP,BP POP BP POP DI POP SI POP CX RET .SECT .DATA mesg1: .ASCIZ "Have a look\n" mesg2: .ASCIZ "qrst\n" .SECT .BSS (a)

! 1 ! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13 ! 14 ! 15 ! 16 ! 17 ! 18 ! 19 ! 20 ! 21 ! 22 ! 23 ! 24 ! 25 ! 26 ! 27 ! 28 ! 29 ! 30 ! 31 ! 32 ! 33 ! 34

#include "../syscalnr.h"

! 1

start: MOV DI,str PUSH AX MOV BP,SP PUSH "PUTCHAR MOVB AL,’\n’ MOV CX,-1 REPNZ SCASB NEG CX STD DEC CX SUB DI,2 MOV SI,DI 1: LODSB MOV (BP),AX SYS LOOP 1b MOVB (BP),’\n’ SYS PUSH 0 PUSH "EXIT SYS .SECT .DATA str: .ASCIZ "reverse\n"

! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13 ! 14 ! 15 ! 16 ! 17 ! 18 ! 19 ! 20 ! 21 ! 22 ! 23 ! 24

(b)

Figure C-17. (a) Copy a string (strngcpy.s). (b) Print a string backward (reverspr.s).

748

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

C.8.5 String Manipulation and String Instructions The main purpose of this section is to show how to handle repeatable string instructions. In Fig. C-17, there are two simple string manipulation programs, strngcpy.s and reverspr.s, both present in the examples directory. The one in Fig. C-17(a) is a subroutine for copying a string. It calls a subroutine, stringpr, which can also be found in a separate file stringpr.s. It is not listed in this appendix. In order to assemble programs containing subroutines in separate source files, just list all source files in the as88 command, starting with the source file for the main program, which determines the names of the executable and the auxiliary files. For example, for the program of Fig. C-17(a) use as88 strngcpy.s stringpr.s

The program of Fig. C-17(b) outputs strings in reverse order. We will now look at them in turn. To demonstrate that the line numbers are really just comments, in Fig. C17(a) we have numbered the lines starting with the first label, omitting what comes before them. The main program, on lines 2 through 8, first calls strngcpy with two arguments, the source string, mesg2, and the destination string, mesg1, in order to copy the source to the destination. Now let us look at strngcpy, starting on line 9. It expects that the addresses of the destination buffer and the string source have been pushed onto the stack just before the subroutine is called. On lines 10 to 13, the registers used are saved by pushing them onto the stack so that they can be restored later on lines 27 to 30. On line 14, we copy SP to BP in the usual way. Now BP can be used to load the arguments. Again, on line 26, we clean the stack by copying BP to SP. The heart of the subroutine is the instruction REP MOVSB, on line 24. The instruction MOVSB moves the byte pointed to by SI to the memory address pointed to by DI. Both SI and DI are then incremented by 1. The REP creates a loop in which this instruction is repeated, decrementing CX by 1 for each byte moved. The loop is terminated when CX reaches 0. Before we can run the REP MOVSB loop, however, we have to set up the registers, which is done in lines 15 through 22. The source index, SI, is copied from the argument on the stack on line 21; the destination index, DI, is set up on line 22. Obtaining the value of CX is more involved. Note that the end of a string is indicated by a zero byte. The MOVSB instruction does not affect the zero flag, but the instruction SCASB (scan byte string) does. It compares the value pointed to by DI with the value in AL, and it increments DI on the fly. Moreover, it is repeatable like MOVSB. So, on line 15 AX and hence AL is cleared, on line 16 the pointer for DI is fetched from the stack, and CX is initialized to −1 on line 17. On line 18, we have the REPNZ SCASB, which does the comparison in loop context, and sets the zero flag on equality. At each step of the loop, CX is decremented, and the loop stops when the zero flag is set, because the REPNZ checks both the zero flag and

SEC. C.8 CX.

EXAMPLES

749

The number of steps for the MOVSB loop is now computed as the difference of the current value of CX and the previous −1 on lines 19 and 20. It is cumbersome that there are two repeatable instructions necessary, but this is the price for the design choice that move instructions never affect the condition codes. During the loops, the index registers have to be incremented, and to this end it is necessary that the direction flag is cleared. Lines 23 and 25 print the copied string by means of a subroutine, stringpr, which is in the examples directory. It is straightforward and will not be discussed here. In the reverse print program shown in Fig. C-17(b), the first line includes the usual system call numbers. On line 3, a dummy value is pushed onto the stack, and on line 4, the base pointer, BP, is made to point to the current top of stack. The program is going to print ASCII characters one by one, thus the numerical value " PUTCHAR is pushed onto the stack. Note that BP points to the character to be printed when a SYS call is made. Line 2, 6 and 7 prepare the registers DI, AL and CX for the repeatable SCASB instruction. The count register and the destination index are loaded in a similar way as in the string copy routine, but the value of AL is the new line character, instead of the value 0. In this way, the SCASB instruction will compare the characters values of the string str to \n instead of to 0, and set the zero flag whenever it is found. The REP SCASB increments the DI register, so, after a hit, the destination index points at the zero character following the new line. On line 12, DI is decremented by two to have it point to the last letter of the word. If the string is scanned in reverse order and printed character by character, we have obtained our goal, so on line 10 the direction flag is set to reverse the adjustment of the index registers in the string instructions. Now the LODSB on line 14 copies the character in AL, and on line 15 this character is put just next to the " PUTCHAR on the stack, so the SYS instruction prints it. The instructions on lines 18 and 19 print an additional new line and the program closes with an " EXIT call in the usual way. The current version of the program contains a bug. It can be found if the program is traced step by step. The command /str will put the string str in the tracer data field. Since the numerical value of the data address is also given, we can find out how the index registers run through the data with respect to the position of the string. The bug, however, is encountered only after hitting the return many times. By using the tracer commands we can get to the problem faster. Start the tracer and give the command 13 to put us in the middle of the loop. If we now give the command b we set a breakpoint on this line 15. If we give two new lines, then we see that the final letter e is printed in the output field. The r command will keep the tracer running until either a breakpoint or the end of the process is encountered. In this way, we can run through the letters by giving the r command repeatedly

750

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

until we are close to the problem. From this point, we can run the tracer at one step at a time until we see what happens at the critical instructions. We can also put the breakpoint at a specific line, but then we must keep in mind, that the file ../syscalnr is included, which causes the line numbers to be offset by 20. Consequently, the breakpoint on line 16 can be set by the command 36b. This is not an elegant solution, so it is much better to use the global label start on line 2 before the instruction and give the command /start+14b, which puts the breakpoint in the same place without having to keep track of the size of the included file.

C.8.6 Dispatch Tables In several programming languages, there exist case or switch statements to select a jump from several alternatives according to some numerical value of a variable. Sometimes, such multiway branches are also needed in assembly language programs, too. Think, for instance, of a set of system call subroutines combined in a single SYS trap routine. The program jumptbl.s, shown in Fig. C18. shows how such a multi-branch switch can be programmed in 8088 assembler. The program starts by printing the string whose label is strt, inviting the user to type an octal digit (lines 4 through 7). Then a character is read from standard input (lines 8 and 9). If the value is AX is less than 5, the program interprets it as an end of file marker and jumps to the label 8 on line 22 to exit with a status code of 0. If end of file has not been encountered, the incoming character, in AL, is inspected. Any character less than the digit 0 is considered to be white space and is ignored by the jump on line 13, which retrieves another character. Any character over digit 9 is considered to be incorrect input. On line 16, it is mapped onto the ASCII colon character, which is the successor of digit 9 in the ASCII character sequence. Thus, on line 17 we have a value in AX between digit 0 and the colon. This value is copied into BX. On line 18, the AND instruction masks off all but the low-order four bits, which leaves the number between 0 and 10 (due to the fact that ASCII 0 is 0x30). Since we are going to index into a table of words, rather than bytes, the value in BX is multiplied by two using the left shift on line 19. On line 20, we have a call instruction. The effective address is found by adding the value of BX to the numerical value of label tbl, and the contents of this composite address are loaded into the program counter, PC. This program chooses one out of ten subroutines according to a character which is fetched from standard input. Each of those subroutines pushes the address of some message onto the stack and then jumps to a " PRINTF system subroutine call which is shared by all of them.

SEC. C.8 #include "../syscalnr.h" .SECT .TEXT jumpstrt: PUSH strt MOV BP,SP PUSH "PRINTF SYS PUSH "GETCHAR 1: SYS CMP AX,5 JL 8f CMPB AL,’0’ JL 1b CMPB AL,’9’ JLE 2f MOVB AL,’9’+1 2: MOV BX,AX AND BX,0Xf SAL BX,1 CALL tbl(BX) JMP 1b 8: PUSH 0 PUSH "EXIT SYS

751

EXAMPLES ! 1 ! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13 ! 14 ! 15 ! 16 ! 17 ! 18 ! 19 ! 20 ! 21 ! 22 ! 23 ! 24

rout0: MOV AX,mes0 JMP 9f rout1: MOV AX,mes1 JMP 9f rout2: MOV AX,mes2 JMP 9f rout3: MOV AX,mes3 JMP 9f rout4: MOV AX,mes4 JMP 9f rout5: MOV AX,mes5 JMP 9f rout6: MOV AX,mes6 JMP 9f rout7: MOV AX,mes7 JMP 9f rout8: MOV AX,mes8 JMP 9f erout: MOV AX,emes 9: PUSH AX PUSH "PRINTF SYS ADD SP,4 RET

.SECT .DATA tbl: .WORD rout0,rout1,rout2,rout3,rout4,rout5,rout6,rout7,rout8,rout8,erout mes0: .ASCIZ "This is a zero.\n" mes1: .ASCIZ "How about a one.\n" mes2: .ASCIZ "You asked for a two.\n" mes3: .ASCIZ "The digit was a three.\n" mes4: .ASCIZ "You typed a four.\n" mes5: .ASCIZ "You preferred a five.\n" mes6: .ASCIZ "A six was encountered.\n" mes7: .ASCIZ "This is number seven.\n" mes8: .ASCIZ "This digit is not accepted as an octal.\n" emes: .ASCIZ "This is not a digit. Try again.\n" strt: .ASCIZ "Type an octal digit with a return. Stop on end of file.\n"

! 25 ! 26 ! 27 ! 28 ! 29 ! 30 ! 31 ! 32 ! 33 ! 34 ! 35 ! 36 ! 37 ! 38 ! 39 ! 40 ! 41 ! 42 ! 43 ! 44 ! 45 ! 46 ! 47 ! 48 ! 49 ! 50 ! 51 ! 52 ! 53 ! 54 ! 55 ! 56 ! 57 ! 58 ! 59 ! 60 ! 61

Figure C-18. A program demonstrating a multiway branch using a dispatch table.

In order to understand what is happening, we need to be aware that the JMP and CALL instructions load some text segment address in PC. Such an address is just a binary number, and during the assembly process all addresses are replaced by their binary values. Those binary values can be used to initialize an array in the data segment, and this is done in line 50. Thus, the array starting at tbl contains the starting addresses of rout 0, rout 1, rout 2, and so on, two bytes per

752

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

address. The need for 2-byte addresses explains why we needed the 1-bit shift on line 19. A table of this type is often called a dispatch table. How those routines work can be seen in the erout routine on lines 43 through 48. This routine handles the case of an out-of-range digit. First, the address of the message (in AX) is pushed onto the stack on line 43. Then the number of the " PRINTF system call is pushed onto the stack. After that, the system call is made, the stack is cleaned up, and the routine returns. The other nine routines, rout0 through rout8, each load the addresses of their private messages in AX, and then jump to the second line of erout to output the message and finish the subroutine. In order to get accustomed to the dispatch tables, the program should be traced with several different input characters. As an exercise, the program can be changed in such a way that all characters generate a sensible action. For example, all characters other than the octal digits should give an error message.

C.8.7 Buffered and Random File Access The program InFilBuf.s, shown in Fig. C-19, demonstrates random I/O on files. A file is assumed to consist of some number of lines, with different lines potentially having different lengths. The program first reads the file and builds a table in which entry n is the file position at which line n begins. Afterward, a line can be requested, its position looked up in the table, and the line read in by means of lseek and read system calls. The file name is given as the first input line on standard input. This program contains several fairly independent chunks of code, which can be modified for other purposes. The first five lines simply define the system call numbers and the buffer size, and set the base pointer at the top of the stack, as usual. Lines 6 through 13 read the file name from standard input, and store it as a string at label linein. If the file name is not properly closed with a new line, then an error message is generated, and the process exits with a nonzero status. This is done in lines 38 through 45. Note that the address of the file name is pushed on line 39, and the address of an error message is pushed on line 40. If we examine the error message itself, (on line 113) then we have a %s string request in the " PRINTF format. The contents of the string linein are inserted here. If the file name can be copied without problems, the file is opened on lines 14 to 20. If the open call fails, then the return value is negative and a jump is made to the label 9 on line 28 to print an error message. If the system call succeeds, then the return value is a file descriptor, which is stored in the variable fildes. This file descriptor is needed in the subsequent read and lseek calls. Next, we read the file in blocks of 512 bytes, each of which is stored in the buffer buf. The buffer allocated is two bytes larger than the necessary 512 bytes, just to demonstrate how a symbolic constant and a integer can be mixed in an expression (on line 123). In the same way, on line 21 SI is loaded with the address of the second element of the linh array, which leaves a machine word containing 0

SEC. C.8

753

EXAMPLES

#include "../syscalnr.h" bufsiz = 512 .SECT .TEXT infbufst: MOV BP,SP MOV DI,linein PUSH "GETCHAR 1: SYS CMPB AL,’\n’ JL 9f JE 1f STOSB JMP 1b 1: PUSH 0 PUSH linein PUSH "OPEN SYS CMP AX,0 JL 9f MOV (fildes),AX MOV SI,linh+2 MOV BX,0 1: CALL fillbuf CMP CX,0 JLE 3f 2: MOVB AL,’\n’ REPNE SCASB JNE 1b INC (count) MOV AX,BX SUB AX,CX XCHG SI,DI STOS XCHG SI,DI CMP CX,0 JNE 2b JMP 1b 9: MOV SP,BP PUSH linein PUSH errmess PUSH "PRINTF SYS

! 1 ! 2 ! 3 ! 4 ! 5 ! 6 ! 7 ! 8 ! 9 ! 10 ! 11 ! 12 ! 13 ! 14 ! 15 ! 16 ! 17 ! 18 ! 19 ! 20 ! 21 ! 22 ! 23 ! 24 ! 25 ! 26 ! 27 ! 28 ! 29 ! 30 ! 31 ! 32 ! 33 ! 34 ! 35 ! 36 ! 37 ! 38 ! 39 ! 40 ! 41 ! 42

PUSH "EXIT PUSH "EXIT SYS 3: CALL getnum CMP AX,0 JLE 8f MOV BX,(curlin) CMP BX,0 JLE 7f CMP BX,(count) JG 7f SHL BX,1 MOV AX,linh-2(BX) MOV CX,linh(BX) PUSH 0 PUSH 0 PUSH AX PUSH (fildes) PUSH "LSEEK SYS SUB CX,AX PUSH CX PUSH buf PUSH (fildes) PUSH "READ SYS ADD SP,4 PUSH 1 PUSH "WRITE SYS ADD SP,14 JMP 3b 8: PUSH scanerr PUSH "PRINTF SYS ADD SP,4 JMP 3b 7: PUSH 0 PUSH "EXIT SYS fillbuf: PUSH bufsiz

! 43 ! 44 ! 45 ! 46 ! 47 ! 48 ! 49 ! 50 ! 51 ! 52 ! 53 ! 54 ! 55 ! 56 ! 57 ! 58 ! 59 ! 60 ! 61 ! 62 ! 63 ! 64 ! 65 ! 66 ! 67 ! 68 ! 69 ! 70 ! 71 ! 72 ! 73 ! 74 ! 75 ! 76 ! 77 ! 78 ! 79 ! 80 ! 81 ! 82 ! 83 ! 84

PUSH buf PUSH (fildes) PUSH "READ SYS ADD SP,8 MOV CX,AX ADD BX,CX MOV DI,buf RET

! 85 ! 86 ! 87 ! 88 ! 89 ! 90 ! 91 ! 92 ! 93

getnum: MOV DI,linein PUSH "GETCHAR 1: SYS CMPB AL,’\n’ JL 9b JE 1f STOSB JMP 1b 1: MOVB (DI),’ ’ PUSH curlin PUSH numfmt PUSH linein PUSH "SSCANF SYS ADD SP,10 RET

! 94 ! 95 ! 96 ! 97 ! 98 ! 99 !100 !101 !102 !103 !104 !105 !106 !107 !108 !109 !110

.SECT .DATA !111 errmess: !112 .ASCIZ "Open %s failed\n" !113 numfmt: .ASCIZ "%d" !114 scanerr: !115 .ASCIZ "Type a number.\n"!116 .ALIGN 2 !117 !118 .SECT .BSS linein: .SPACE 80 !119 fildes: .SPACE 2 !120 linh: .SPACE 8192 !121 curlin: .SPACE 4 !122 buf: .SPACE bufsiz+2 !123 !124 count: .SPACE 2

Figure C-19. A program with buffered read and random file access.

at the bottom of this array. The register BX will contain the file address of the first unread character of the file, and hence, it is initialized to 0 before the first time that the buffer is filled on line 22.

754

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

The filling of the buffer is handled by the fillbuf routine on lines 83 through 93. After pushing the arguments for the read, the system call is requested, which puts the number of characters actually read in AX. This number is copied into CX and the number of characters still in the buffer will be kept in CX thereafter. The file position of the first unread character in the file is kept in BX, so CX has to be added to BX in line 91. On line 92, the buffer bottom is put into DI in order to get ready to scan the buffer for the next new line character. After returning from fillbuf, line 24 checks whether anything was actually read. If not, we jump out of the buffered read loop to the second part of the program in line 25. Now we are ready to scan the buffer. The symbol \n is loaded into AL on line 26, and in line 27 this value is scanned for by REP SCASB loop compared to the symbols in the buffer. There are two ways to exit the loop: either when CX hits zero or when a scanned symbol is a new line character. If the zero flag is set, then the last symbol scanned was a \n and the file position of the current symbol (one after the new line), is to be stored in the linh array. The count is then incremented and the file position is computed from BX and the number of characters still available is in CX (lines 29 through 31). Lines 32 through 34 perform the actual store, but since STOS assumes that the destination is in DI instead of in SI, these registers are exchanged before and after the STOS. Lines 35 through 37 check whether more data is available in the buffer, and jump according to the value of CX. When the end of the file is reached, we have a complete list of file positions of the heads of the lines. Because we started the linh array with a 0 word, we know that the first line started at address 0, and that the next line starts at the position given by linh + 2 etc. The size of line n can be found from the starting address of line n + 1 minus the start address of line n. The aim of the rest of the program is to read the number of a line, to read that line into the buffer, and to output it by means of a write call. All the necessary information can be found in the linh array, whose nth entry contains the position of the start of line n in the file. If the line number requested is either 0 or out of range, the program exits by jumping to label 7. This part of the program starts with a call to the getnum subroutine on line 46. This routine reads a line from standard input and stores it in the linein buffer, (on lines 95 through 103). Next, we prepare for the SSCANF call. Considering the reverse order of the arguments, we first push the address of curlin, which can hold an integer value, then the address of the integer format string numfmt, and finally the address of the buffer linein containing the number in decimal notation. The system subroutine SSCANF puts the binary value in curlin if possible. On failure, it returns a 0 in AX. This return value is tested on line 48; on failure, the program generates an error message through label 8. If the getnum subroutine returns a valid integer in curlin, then we first copy it in BX. Next we test the value against the range in lines 49 through 53, generating an EXIT if the line number is out of range.

SEC. C.8

EXAMPLES

755

Then we must find the end of the selected line in the file and the number of bytes to be read, so we multiply BX by 2 with a left shift SHL. The file position of the intended line is copied to AX on line 55. The file position of the next line is squirreled away in CX and will be used to compute the number of bytes in the current line. To do a random read from a file, an lseek call is needed to set the file offset to the byte to be read next. The lseek is performed with respect to the start of the file, so first an argument of 0 is pushed to indicate this on line 57. The next argument is the file offset. By definition, this argument is a long (i.e., 32-bit) integer, so we first push a 0 word and then the value of AX on lines 58 and 59 to form a 32-bit integer. Then the file descriptor and code for LSEEK are pushed and the call is made on line 62. The return value of LSEEK is the current position in the file and can be found in the DX : AX register combination. If the number fits into a machine word (which it will for files shorter than 65536 bytes), then AX contains the address, so subtracting this register from CX (line 63), yields the number of bytes to be read in order to bring the line into the buffer. The rest of the program is easy. The line is read from the file on lines 64 through 68 and then it is written to standard output via file descriptor 1 on lines 70 through 72. Note that the count and the buffer value are still on the stack after the partial stack cleanup on line 69. Finally, on line 73, we reset the stack pointer completely and we are ready for the next step, so we jump back to label 3, and restart with another call to getnum. Acknowledgements The assembler used in this appendix is part of the ‘‘Amsterdam Compiler Kit.’’ The full kit is available online at www.cs.vu.nl/ack. We thank the people who were involved in the original design: Johan Stevenson, Hans Schaminee, and Hans de Vries. We are especially indebted to Ceriel Jacobs, who maintains this software package, and who has helped adapt it several times to meet the teaching requirements of our classes, and also to Elth Ogston for reading the manuscript and testing the examples and exercises. We also want to thank Robbert van Renesse and Jan-Mark Wams, who designed tracers for the PDP-11 and the Motorola 68000, respectively. Many of their ideas are used in the design of this tracer. Moreover, we wish to thank the large group of teaching assistants and system operators who have assisted us during many assembly language programming courses over a period of many years.

756

ASSEMBLY LANGUAGE PROGRAMMING

APP. C

PROBLEMS 1. After the instruction MOV AX, 702 is executed, what are the decimal values for the contents of AH and AL? 2. The CS register has the value 4. What is the range of absolute memory addresses for the code segment? 3. What is the highest memory address the 8088 can access? 4. Suppose that CS = 40, DS = 8000, and IP = 20. a. What is the absolute address of the next instruction? b. If MOV AX, (2) is executed, which memory word is loaded into AX? 5. A subroutine with three integer arguments is called following the calling sequence described in the text, that is, the caller pushes the arguments onto the stack in reverse order, then executes a CALL instruction. The callee then saves the old BP and sets the new BP to point to the saved old one. Then the stack pointer is decremented to allocate space for local variables. With these conventions, give the instruction needed to move the first argument into AX. 6. In Fig. C-1 the expression de − hw is used as an operand. This value is the difference of two labels. Might there be circumstances in which de + hw could be used as a valid operand? Discuss your answer. 7. Give the assembly code for computing the expression: x=a+b+2

8. A C function is called by foobar(x, y);

Give the assembly code for making this call. 9. Write an assembly language program to accept input expressions consisting of an integer, an operator, and another integer and output the value of the expression. Allow the +, −, ×, and / operators.

INDEX

Numbers

Adder (continued) half, 153-155 ripple carry, 155 Additive inverse, 379 Address, 70, 704 memory, 70-71 Address decoding, 222-225 Address space, 429 Addressing, 351, 361-362 8051, 373-374 8088, 711-714 based-indexed, 365 branch instruction direct, 361 displacement, 713 immediate, 361 indirect, 362-363, 712 indexed, 363-365 Pentium 4, 371-373 register indirect, 362-363 register, 361-362 stack, 365-368 UltraSPARC, 373

8051 (see Intel 8051) 8088 (see Intel 8088)

A Absolute path, 486 Accelerated graphics port 110, 206 Access control list, 492 Access token, 491 Accumulator, 18, 323, 347 Accumulator register, 705 ACK (see Amsterdam Compiler Kit) Acknowledgment packet, 216 ACL (see Access Control List) Active matrix display, 109 Actual parameter, 520 Adder, 153-155 carry select, 155 full 154-155 757

758

INDEX

Addressing mode, 360-375 8051, 373-374 based-indexed, 365 branch instruction, 369 direct, 361 discussion, 374-375 immediate, 361 indexed, 363-365 Pentium 4, 371-372 register, 361-362 stack, 365-368 UltraSPARC III, 373 ADSL (see Asymmetric DSL) Advanced microcontroller bus architecture, 567 Advanced programmable interrupt controller, 193 Aggregate bandwidth, 644 AGP bus, 110, 206-207 Aiken, Howard, 16 Algorithm, 8 Allocation/renaming unit, Pentium 4, 315 Alpha, 25 Alpha channel mask, 579 ALU (see Arithmetic Logic Unit) AMBA (see Advanced Microcontroller Bus Architecture) Amdahl’s law, 645 American Standard Code for Information Interchange, 128-129 Amplitude modulation, 118 Amsterdam compiler kit, 727 Analytical engine, 15 APIC (see Advanced Programmable Interrupt Controller) Apple II, 24 Apple Macintosh, 25 Application layer, 650 Application programming interface, 478 Application-specific integrated circuit, 571 Arbitratio, PCI bus, 208-209 Architecture, 8, 56 Argument, 721 Arithmetic, binary, 688 Arithmetic circuit, 152-157 Arithmetic Logic Unit, 6, 52-54, 156-157

Array processor, 66-67 ASCII (see American Standard Code for Information Interchange) ASIC (see Application-Specific Integrated Circuit) Assembler directive, 515 Assembler, 7, 508, 522-530, 702 8088, 725-731 pass one, 523-527 pass two, 527-529 symbol table, 529-530 two-pass, 622-523 Assembly language, 508-522, 701-704 8051, 725-731 characteristics, 508-509 pseudoinstructions, 515-517 statements, 512-515 why use, 509-512 Assembly language level, 507-546 Assembly language program, 726 Assembly language programming, 701-756 overview, 702-704 Assembly process, 522-530 Asserted signal, 168 Associative memory, 445, 529 Asymmetric DSL, 120 Asynchronous bus, 182-184 AT Attachment, 87 ATA packet interface, 87 ATA-3, 87 Atanasoff, John, 16, 18 ATAPI-4, 87 Attraction memory, 611 Auto decrement, 714 Auto increment, 714 Auxiliary carry flag, 718

B Babbage, Charles, 15-16 Baby feeding algorithms, 435 Backward compatibity, 332 Bardeen, John, 19

INDEX Base, 136 Base pointer, 708 Base register, 706 Based-indexed addressing, 365 Basic block, 309 Basic input output system, 86 Batch system, 11 Baud, 119 Bayer filter, 125 BCD (see Binary coded decimal) Bechtolsheim, Andy, 42 Benchmark, 510 Best fit algorithm, 444 Big endian machine, 71 Binary arithmetic, 688 Binary Coded Decimal, 69, 712 Binary number, 679-690 addition of, 688 conversion between radices, 681-685 negative, 685-688 Binary program, 702 Binary search, 529 Binding time, 536, 538 BIOS (see Basic Input Output System) Bipolar, 138 Bisection bandwidth, 614 Bit, 69-70, 682 Bit map, 349 Bit slice, 157 Bitblt, 580 Block cache, 473 Block started by symbol, 727 Block transfer, bus, 186-189 Blocking network, 602 Blu-Ray, 102 BlueGene/L, 618-622, 626-627 Boole, George, 138 Boolean algebra, 138-140 Branch, 395-396 Branch history shift register, 303 Branch prediction, 299-304 dynamic, 301-303 static, 303-304 Branch table, 318 Branch target buffer, 314 Brattain, Walter, 19

759 Breakpoint, 735 Broadband, 120 BSS (see Block Started by Symbol) Bubblejet, 113 Bucket, 529 Buffer, circular, 462 Buffered message passing, 633 Bundle, 417 Burroughs B5000, 21 Bus, 20, 51, 102-105, 176-195, 202-220 AGP, 206-207 asynchronous, 182-184 block transfer, 186-189 EISA, 204 handshaking, 184 IBM PC, 203 ISA, 203-204 multiplexed, 179 PCI, 204-212 PCI Express, 212-217 Pentium 4, 194-195 protocol, 177 synchronous, 180-182 USB, 217-220 Bus arbiter, 104 Bus arbitration, 184-187 PCI, 208-209 Bus clocking, 180-182 Bus cycle, 180 Bus driver, 177 Bus grant, 184-185 Bus master, 177 Bus operation, 187-189 Bus protocol, 177 Bus receiver, 178 Bus skew, 179 Bus slave, 177 Bus timing, 180-182 Bus transceiver, 178 Bus width, 178-179 Busy waiting, 385 Byron, Lord, 15 Byte, 71, 335, 712 Byte instruction, 711 Byte ordering, 71-73 Byte register, 711

760

C

INDEX

Cell, memory, 70 Central processing unit, 18, 51 Character code, 127-131 Cable Internet, 122-125 Cache memory, 38-39, 77-80, 293-298, 595-599 Charge-coupled device, 125 Checkerboarding, 443 direct-mapped, 295-297 Child process, 494 invalidate strategy, 597 Chip, 146 level 2, 293-294 CPU, 173-176 MESI protocol, 597-599 Circuit multiple levels, 293 arithmetic, 152-157 set-associative, 297-298 combinational, 147-152 snooping, 595-597 Circuit equivalence, 141-145 split, 80, 293 Clock, 157-159 unified, 80 Clock cycle time, 157 update strategy, 597 Clocked D latch, 161 write-allocate, 298, 597 Clocked SR latch, 160-161 write-back, 298, 597 Clone, 24 write-deferred, 298 Closure, 680 write-through, 298 Cluster Cache coherence protocol, 595 Google, 628-632 MESI, 597-599 NTFS, 492 Cache coherent NUMA multiprocessors, workstation, 34-36, 589, 627-632 603-611 Cluster of workstations, 34-36, 589 Cache consistency, 595 CLUT (see Color Look Up Table) Cache hit, 296 Coarse-grained multithreading, 557 Cache line, 79, 294, 595 COBOL program, 36 Cache miss, 296 Code generation, 726 Cache only memory access multiprocessors, Code page, 128 589, 611-612 Code point, 129 Call gate, 449 Code segment, 704 Camera, digital, 125-127 Codesign, 26 Carrier, 118 Codeword, 73 Carry select adder, 155 Collective layer, 650 Catamount, 625 Collector, 136 Cathode Ray Tube, 106-107 Color gamut, 116 CC-NUMA (see Cache Coherent NUMA Color look up table, 579 multiprocessors) Color palette, 109 CCD (see Charge-Coupled Device) COLOSSUS, 17 CD-R, 97 COMA (see Cache Only Memory Access CD-recordable, 97 multiprocessors) CD-rewritable, 99 Combinational circuit, 147-152 CD-ROM (see Compact Disc-Read comparator, 149-152 Only Memory) decoder, 148-149, 151 CDC (see Control Data Corporation) multiplexer, 148 CDC 6600, 14, 21, 64, 305, 567 Committed page, 480 CDC Cyber, 57 Commodity off the shelf, 35 Celeron, 39

761

INDEX Communicator, 634 Compact disc-read only memory, 93-97 sector, 95 track, 97 Rewritable, 99 XA, 97 Comparator, 149-150 Comparison and branch instructions, 379-381 Comparison of architectures, 325-326 Compiler, 7, 508 Complex instruction set computer, 58 Computer architecture, 8 milestones, 13-26 Computer center, 24 Computer zoo, 27-36 Condition code, 338 Condition code register, 709 Condition variable, 496 Conditional execution, 418 Consistency model (see Memory semantics) Constant pool, 248 Constant pool pointer, 233, 242, 249, 260 Consumer, 462 Context, 451 Control Data Corporation, 21 Control signal, 237 Control store, 57, 241 Controller, 102 Conversion between radices, 683-685 Coprocessor, 567-582 Copy on write, 480 Core, 562 Core dump, 10 CoreConnect, 565 Coroutine, 401-404 COTS (see Commodity Off The Shelf) Counter register, 707 COW (see Cluster Of Workstations) CP/M, 24 CPP (see Constant Pool Pointer) CPU (see Central Processing Unit) CPU chip, 174-176 Cray, Seymour, 21 Cray-1, 21, 57 CRC (see Cyclic Redundancy Check) Critical section, 499

Crossbar switch, 599-600 Crosspoint, 599 CRT (see Cathode Ray Tube) Cryptoprocessor, 581-582 Cube, 617 Cycle stealing, 104, 386 Cyclic redundancy check, 215, 220, 570 Cylinder, 84 CYMK printer, 116

D D latch, 161 Daisy chaining, 185 Data cache, 319 Data movement instruction, 375-376 Data path, 6, 52, 232-246 Mic-1, 242 Mic-2, 280 Mic-3, 284 Mic-4, 289 timing, 235-237 Data path cycle, 54 Data register, 707 DATA section, 727 Data segment, 719 Data type, 348-351 nonnumeric, 348-349 numeric, 349-350 DDR (see Double Data Rate memory) DEC (see Digital Equipment Corporation) DEC Alpha, 25 DEC PDP-1, 20 DEC PDP-8, 20 DEC VAX, 57, 58 Decimal, 728 Decoder, 148-149 Decoding unit, 289 Degree, 614 Delay slot, 300 Demand paging, 435 DeMorgan’s law, 143-144 Demultiplexer, 148 Denormalized number, 697

762

INDEX

Design principles, RISC, 59-65 Destination index, 708 Destination operand, 711 Device driver, 476 Device level, 5, 136 Device register bus, 566 Diameter, network, 614 Dibit modulation, 119 Difference engine, 15 Digital camera, 125-127 Digital Equipment Corporation, 20, 21, 57 Digital logic level, 5, 135-230 buses, 176-220 CPU chips, 173-176 circuits, 146-159 gates, 135-145 I/O interfacing, 221-225 memory, 159-173 Digital subscriber line, 119-122 Digital subscriber line access multiplexer, 122 Digital versatile disk, 99-102 Digital video disk, 100 Dimensionality, 616 DIMM (see Dual Inline Memory Module) DIP (see Dual Inline Package) Direct addressing, 361, 712 Direct memory access, 103, 385 Direct-mapped cache, 295 Direction flag, 714 Directory, 459-460 Directory-based multiprocessor, 606-611 Disk, 81-102 CD-ROM, 93-97 DVD, 99-102 IDE, 86-88 magnetic, 82-93 optical 93-102 RAID, 89-93 SCSI, 88-89 Winchester, 83-84 Disk controller, 85 Diskette, 86 Dispatch table, 752 Distributed memory system, 583 Distributed shared memory, 585, 636-638 DLL (see Dynamic Link Library)

DMA (see Direct Memory Access) Dot, 726 Dots per inch, 113 Double, 707 Double data rate memory, 172 Double indirect block, 488 Double pointer, 325 Double precision, 348 Double torus, 616 DPI (see Dots Per Inch) DRAM (see Dynamic RAM) DSL (see Digital Subscriber Line) DSLAM (see Digital Subscriber Line Access Multiplexer) DSM (see Distributed Shared Memory) Dual, 143 Dual inline memory module, 80l Dual inline package, 146 DVD (see Digital Versatile Disk) Dyadic instruction, 376-377 Dye-based ink, 116 Dye-sublimation printer, 117 Dynamic link library, 539 Dynamic linking, 539-542 Dynamic RAM, 171 Dynamic relocation, 536-538

E Eckert, J. Presper, 17 ECL (see Emitter-Coupled Logic) Edge-triggered flip-flop, 162 EDO (see Extended Data Output memory) EDVAC (see Electronic Discrete Variable Automatic Computer) EEPROM (see Electrically Erasable PROM) Effective address, 714 Egress processing, 573 EHCI (see Enhanced Host Controller Interface) EIDE (see Extended IDE) EISA (see Extended ISA) Electrically erasable PROM, 173 Electronic Discrete Variable Automatic Computer, 17 Electronic Numerical Integrator And Computer, 17 Embedded computer, 26

INDEX Emitter, 136 Emitter-coupled logic, 138 Emulation, 23 Enable, 161 Endian, 71-72 big, 71 little, 72 Enhanced Host Controller Interface, 220 ENIAC (see Electronic Numerical Integrator And Computer) ENIGMA, 16-17 Entry point, 535 Environmental subsystem, Windows XP, 477 EPIC (see Explicitly Parallel Instruction Computing) EPROM (see Erasable PROM) Erasable PROM, 44, 172 Error-correcting code, 73-77 Escape code, 358 Estridge, Philip, 24 Ethernet, 568 Event, 499 Evolution of computers, 8-13 Excess notation, 686 Executable binary file, 726 Executable binary program, 508, 531 Executive, Windows XP, 476 Expanding opcode, 354-357 Explicit linking, 542 Explicitly parallel instruction computing, 414-421 Exponent, 692 Extended data output memory, 171 Extended IDE, 87 Extended ISA, 104, 204 External fragmentation, 443 External reference, 533 External symbol, 535 Extra segment, 719

F Fabric layer, 650 False sharing, 638

763 Fanout, 614 Far call, 721 Far jump, 719 Fast page mode memory, 171 FAT (see File Allocation Table) Fat tree, 616 Fetch-decode-execute cycle, 54 Fetch-execute cycle, 232 Fiber, 498 Field programmable gate array, 571 FIFO (see First-In First-Out algorithm) Fifth generation project, Japanese, 26 File, 454-460 File allocation table, 489 File cache manager, Windows XP, 476 File descriptor, 483, 724 File index, 456 File system UNIX, 482-489 Windows XP, 489-493 Filter, 484 Fine-grained multithreading, 556 Finite state machine Branch prediction, 302-303 Instruction fetch unit, 278-279 Finite-precision number, 679-683 Firewall, 569 First fit algorithm, 444 First-generation computer, 16-19 First pass, 726 First-in first-out algorithm, 437 Flags register, 709, 338 Flash memory, 173 Flat panel display, 107-109 Flip-flop, 161-163 Floating-point number, 691-699 Floppy disk, 86 Flow control, 216 Flow of control, ISA level 395-408 branches, 395-396 coroutines, 401-404 interrupts, 404-408 procedures, 396-401 traps, 404 Flynn’s taxonomy, 587-588 Formal parameter, 520

764

INDEX

Forrester, Jay, 19 FORTRAN, 9-10 FORTRAN monitor system, 10-11 Forward reference problem, 523 Fourth-generation computers, 23-26 FPGA (see Field Programmable Gate Array) FPM (see Fast page Mode memory) Fraction, 692 Fragmentation external, 443 internal, 438-439 Frame, 95 Frame pointer, 340 Free list, 457 Free page, 480 Frequency modulation, 118 Frequency shift keying, 118 FSM (see Finite State Machine) Full adder, 154 Full handshake, 184 Full interconnect, 616 Full resource sharing, 560 Full-duplex line, 119 Functional unit, 65

G Game computer, 33 Gamut, 116 Gate, 5, 136-138 Gate delay, 147 GDT (see Global Descriptor Table) General register, 705 Global descriptor table, 445-446 Global label, 727 Globe, 642-643 Goldstine, Herman, 18 Google cluster, 628-632 Graphical user interface, 25, 473 Graphics device interface, Windows XP, 477 Green Book, 96 Grid, 616, 649-651 GUI (see Graphical User Interface)

H H register, 233, 242 Half adder, 153-154 Half-duplex line, 119 Halftone screen frequency, 115 Halftoning, 115 Hamming, Richard, 29 Hamming code, 76-77 Hamming distance, 73 Handle, 478 Hard disk (see Disk) Hardware, 8 equivalence with software, 8 Hardware abstraction layer, 475 Hardware DSM, 603 Harvard architecture, 80 Hash coding, 529 Hashing, 529 Hazard, 286 HDTV (see High Definition TeleVision) Headend, 122 Header, 213 Headless workstation, 627 Hexadecimal, 681, 728 High definition television, 580 High Sierra, 96 High-level language, 7 compared to assembly lanaguage, 509, 512 History 1945-1955, 16-19 1955-1965, 19-21 1965-1980, 22-23 1980-present, 23-26 computer system, 13-26 Intel, 37-42 Sun Mirosystems, 42-44 Hit ratio, 78 Hoagland, Al, 28 Hoff, Ted, 37 Hoisting, code, 310 Host library, 542 HTTP (see HyperText Transfer Protocol) Hypercube, 617 Hypertext transfer protocol, 569 Hyperthreading, 559

INDEX

I I-node, 487 I/O (see Input/Output) I/O instructions, 83-386 I/O manager, Windows XP, 476 IA-32, 339 IA-64, 411-421 bundle, 417 EPIC model, 415 instruction scheduling, 416-418 predication, 418-410 speculative loads, 420-421 IAS machine, 18 IBM 360, 22-23, 26, 567 IBM 701, 19 IBM 704, 19 IBM 801, 58 IBM 1401, 20, 22 IBM 7094, 20, 26 IBM Corporation, 19, 20-23, 203 IBM PC, 24, 28 bus, 203 origin, 24 IBM PS/2, 203 IC (see Integrated Circuit) IDE (see Integrated Drive Electronics) IEEE floating-point standard, 694-698 IFU (see Instruction Fetch Unit) IJVM, 231-232, 246-255 constant pool, 248 data path, 232-235 instruction set, 250-256 Java code, 254-255 local variable frame, 248 memory model, 248-250 memory operation, 237-238 method area, 249 Mic-1 implementation, 260-271 Mic-2 implementation, 271-281 Mic-3 implelentation, 281-288 Mic-4 implementation, 288-292 operand stack, 249 stack, 246-248 timing 235-237 ILC (see Instruction Location Counter)

765 ILLIAC, 17 ILLIAC IV, 66-67, 587 Immediate addressing, 361, 714 Immediate file, 493 Immediate operand, 361 Implicit linking, 542 Implied addressing, 714 Import library, 542 Indexed addressing, 363-365 Indexed color, 109, 579 Indirect block, 488 Industry standard architecture, 104, 203 Infix notation, 365 Informative information, in standard, 334 Ingress processing, 573 Initiator, PCI bus, 207 Inkjet printer, 113 Input/output, 102-131 Input/output devices CRT monitor, 106-107 digital cameras, 125-127 flat panel display, 107-109 keyboards, 105-106 magnetic disks, 82-93 mice, 110-112 modems, 117-119 optical disks, 93-102 printers, 112-117 telecommunications equipment, 119-125 terminals, 105-110 Input/output instructions, 383-386 Instruction, 339, 375-395 8051, 392-393 branch, 379-381, 395-396 comparison, 379-381 data movement, 375-377 dyadic, 376-377 I/O, 383-386 loop, 382-383 monadic, 377-379 Pentium 4, 386-389 procedure call, 381 UltraSPARC, 389-392 Instruction count, relation to RISC, 58-59 Instruction execution,54-56 Instruction fetch unit, 276-280

766

INDEX

Instruction format, 351-375 8051, 359-360 design criteria, 352-354 ISA level, 352-360 Pentium 4, 357-358 UltraSPARC, 358-359 Instruction group, 416 Instruction issue unit, 318 Instruction-level parallelism, 61-65, 549-556 Instruction location counter, 523 Instruction pointer, 704, 708 Instruction register, 52, 323 Instruction scheduling, 416-418 Instruction set, 8088, 715-725 Instruction set architecture, 6, 701 8051, 345-347 Pentium 4, 339-341 UltraSPARC, 341-344 Instruction set architecture level, 6 Instruction sets, comparison, 392-395 Instruction types, 375-395 Integer unit, 43 Integrated circuit, 22, 146-147 Integrated drive electronics, 86 Intel 4004, 37-38 Intel 8008, 37 Intel 8080, 37-38 Intel 8086, 37-38 Intel 8051, 200-202 addressing, 373-374 data types, 351 history, 44-46 instruction formats, 359-360 instructions, 392-393 microarchitecture, 323-325 overview of the ISA, 345-347 Intel 8086, 37-38 Intel 8088, 25, 37-38, 339, 704-709 addressing, 711-714 assembler, 725-731 example programs, 736-755 instruction set, 715-725 memory organization, 709-714 tracer, 732-735 Intel 8255A, 221-222 Intel 80286, 38

Intel 80386, 38 Intel 80486, 38-39 Intel Celeron, 39 Intel Corporation, 37 Intel Pentium 4 (see Pentium 4) Intel Xeon, 40 Interconnection networks, 614-617 bisection bandwidth, 614-615 topology, 614-616 Interlaced display, 580 Interleaved display, 602 Internal fragmentation, 438 Internet over cable, 122-125 Internet protocol, 570 Internet service provider, 569 Interpretation, 2-4 Interpreter, 2, 54-55 Interpreter, 703 Interrupt, 103, 404-408 imprecise, 307 precise, 307 transparent, 406 Interrupt handler, 104 Interrupt service routine, 406-408 Interrupt vector, 189, 405 Intersector gap, 83 Invalidate strategy, 597 Inversion bubble, 137 Inverter, 137 Inverting buffer, 167 Invisible computer, 26 IP header, 570 IP protocol, 570 IR (see Instruction Register) Iron oxide valley, 28 ISA (see Industry Standard Architecture) ISA (see Instruction Set Architecture) ISA bus, 203-204 ISA level, 6, 331-426 overview, 333-347 data types, 348-351 instruction formats, 351-360 addressing, 360-375 instruction types, 375-395 flow of control, 395-411 IA-64, 411-421

INDEX ISP (see Internet Service Provider) Itanium 2, 411-421 IU (see Integer Unit)

J Java, 7 Java Virtual Machine, 231 Jobs, Steve, 24 JOHNIAC, 17 Joint photographic experts group, 127 Joint test action group, 580 Joy, Bill, 42 JPEG (see Joint Photographic Experts Group) JTAG (see Joint Test Action Group) Jump (see Branch) JVM (see Java Virtual Machine) JVM (see also IJVM)

K Kernel, 475 Kernel mode, 335 Key, 456 Keyboard, 105-106 Khosla, Vinod, 42 Kildall, Gary, 24

L L1 BTB, 314 Label, 702 LAN (see Local-Area Network) Land, 93 Lane, 215 Language, 1 Laser printer, 113 Latch, 159-161 Latency, 63 rotational, 84

767 Latency hiding, 648-649 Latin-1, 128 Layer, 3 LBA (see Logical Block Addressing) LCD (see Liquid Crystal Display) LDT (see Local Descriptor Table) Least recently used algorithm, 298, 436 LED (see Light Emitting Diode) Left value, 711 Legacy, 191 Leibniz, Gottfried, 15 Level, 2-8 Level 2 cache, 293 Level-triggered latch, 162 Light Emitting Diode, 111 Linda, 638-640 Linear address, 447 Lines per inch, 115 Link, 486 Link layer, 215 Linkage editor, 531 Linkage segment, 539 Linker, 531-542, 726 Linking binding time, 536-538 dynamic, 539-542 MULTICS, 539 object modules, 535-536 tasks performed, 530-534 UNIX, 542 Windows, 539-541 Liquid crystal display, 107-109 Literal, 525 Little endian machine, 71, 712 Load/store architecture, 344 Loading, 530-545 Local descriptor table, 445-446 Local label, 727 Local loop, 120 Local variable frame, 246, 248 Local variable pointer, 233, 242, 246, 260 Local-area network, 568 Locality principle, 78, 436 Location counter, 726 Logical block addressing, 87 Logical record, 455

768

INDEX

Long, 707, 712 Loop control instruction, 382-383 Loosely coupled system, 548 Lovelace, Ada, 15 LPI (see Lines Per Inch) LRU (see Least Recent Used algorithm) LV (see Local Variable pointer)

M Machine language, 1, 701 Macintosh, Apple, 25 Macro, 517-522 implementation, 521-522 Macro call, 519 Macro definition, 518 Macro expansion, 519 Macro parameter, 520 Macroarchitecture, 246 Magnetic disk, 82-93 Mailslot, 498 Mainframe, 36 MANIAC, 17 Mantissa, 692 MAL (see Micro Assembly Language) MAR (see Memory Address Register) Mark I, 16 Mask, 376 MASM, 512 Massively parallel processor, 589, 617-627 Master, bus, 177 Master file table, 492 Matrix printer, 112 Mauchley, John 16-18 MBR (see Memory Buffer Register) McNealy, Scott, 42 MCS-51 family, 44 MDR (see Memory Data Register) Media processor, 576-581 Memory, 159-173 8088, 709-714 associative, 445, 529 attraction, 611

Memory (continued) cache, 38, 77-80, 293-298, 595-599 DDR, 172 DRAM, 171 dynamic RAM, 171 EDO, 171 EEPROM, 173 EPROM, 172 flash, 173 FPM, 171 primary, 69-81 PROM, 172 SDRAM, 172 static RAM, 171 secondary, 81-102 SRAM, 171 virtual, 429-453 Memory address, 70-71 Memory address register, 233, 237, 242 Memory buffer register, 233, 239, 242 Memory chip, 168-171 Memory data register, 233, 237, 242 Memory hierarchy, 81-82 Memory management unit, 433 Memory map, 430 Memory-mapped I/O, 222 Memory model, 335-337 Memory organization, 164-168 Memory packaging, 80-81 Memory refresh, 16, 171 Memory semantics, 590-594 processor consistency, 592 release consistency, 593-594 sequential consistency, 590-591 weak consistency, 592-593 Mesh network, 616 MESI cache coherence protocol, 597-599 Message passing, 632-633 Message-passing interface, 633 Message-passing multicomputers, 612-649 Message queue, 494 Metal oxide semiconductor, 138 Method, 381 Method area, 249 Metric units, 46-48 MFT (see Master File Table)

INDEX

769

Mic-1, 241, 260-271 Miss ratio, 79 data path, 242 MMU (see Memory Management Unit) implementation, 260-271 MMX (see MultiMedia eXtensions) Mic-2, 281-288 Mnemonic, 702, 726 data path, 280 Modem, 117-119 implementation, 281-284 Modulation, 118 Mic-3, 284 amplitude, 118 data path, 284 frequency, 118 implementation, 284-288 phase, 119 pipeline Moore, Gordon, 27 Mic-4, 288-292 Moore’s law, 27-28, 41 data path, 289 MOS (see Metal Oxide Semiconductor) implementation, 288-292 Motherboard, 102 Mickey, 111 Motif, 473 Micro assemby language, 256 Motion picture experts group, 564 Micro-operation, 290 Motorola 68000, 57 Microarchitecture Mouse, 110-112 8051, 323-325 MPC (see MicroProgram Counter) Pentium 4, 312-317 MPEG-2, 564 UltraSPARC, 317-323 MPI (see Message-Passing Interface) Microarchitecture level, 6, 231-330 MPP (see Massively Parallel Processor) branch prediction, 299-304 MS-DOS, 25 cache memory, 292-298 Multicomputer, 68-69, 583-587, 612-649 design, 271-292 MPP, 617-618 examples, 311-325 BlueGene/L, 618-622, 627 IJVM example, 231-292 Red Storm, 622-627 Microcontroller, 31-33 Google cluster, 628-632 Multicomputer performance, 643-649 Microdrive, 127 Multicomputer scheduling, 635-636 Microinstruction, 58, 239-241 Multicomputer software, 632-635 notation, 255-260 MULTICS (see MULTiplexed Information and Microinstruction control, 241-246 Computing Service) Microinstruction register, 243 Multilevel machine, 4, 8-13 Microprogram, 6, 705 Multimedia extensions, 39 Microprogram counter, 243 Multiple instruction stream multiple data Microprogramming, 9 stream computer, 587-588 history, 12-13, 25 Multiple instruction stream single data Microsoft Corporation, 25, 473-474 stream computer, 587-588 Microstep, 285 Multiplexed bus, 179 Milestones in computer architecture, 13-26 MIMD (see Multiple Instruction-stream Multiple Multiplexed information and computing service, 444-445, 539 Data-stream) Multiplexer, 148 MIPS (acronym), 60 Multiprocessor, 67-68, 320, 582-612 MIPS (chip), 58 bus-based 595-599 MIR (see MicroInstruction Register) MISD (see Multiple instruction stream single data crossbar-based, 599-600 switching network, 600-602 stream computer)

770

INDEX

Multiprocessor on a chip, 562-567 Multiprogramming, 22 Multisession, 98 Multistage switching network, 601 Multithreading, 556-562 Mutex, 496 Mutual exclusion, 496 Myhrvold, Nathan, 28

N N-way set-associative cache, 297 NaN (see Not a Number) Nathan’s first law of software, 28 NC-NUMA (see NonCoherent NUMA multiprocessors) Near call, 721 Near jump, 719 Negated signal, 169 Negative logic, 145 Netburst microarchitecture, 312-317 Network interface device, 121 Network of workstations, 589 Network processor, 568-575 Nexperia, 576-581 Nibble, 387 NID (see Network Interface Device) No remote memory access computer, 589 Nonblocking message passing, 633 Nonblocking network, 599 Noncoherent NUMA multiprocessors, 603-604 Noninverting buffer, 167 Nonuniform Memory Access, 589 Nonuniform memory access multiprocessor, 602-611 Nonvolatile memory, 172-173 NORMA (see NO Remote Memory Access computer) Normalized number, 694 Normative information, in standard, 334 Not a Number, 698 NOW (see Network of Workstations) NT file system, 489 NTFS (see NT File System)

NUMA (see NonUniform Memory Access multiprocessor)

O Object file, 726 Object manager, Windows XP, 476 Object module, 535-536 Object program, 508 OCP-IP (see Open Core Protocol-International Partnership) Octal number, 681, 728 Off-line, 459 OGSA (see Open Grid Services Architecture) OHCI (see Open Host Controller Interface) Omega network, 601 Omnibus, PDP-8, 20 On-chip multithreading, 556-562 On-chip parallelism, 548-567 On-line, 459 One’s complement, 685 OPC (see Opcode register) Opcode, 232 Opcode register, 233, 242, 260 Open collector, 178 Open Core Protocol-International Partnership, 567 Open grid services architecture, 651 Open Host Controller Interface, 220 Operand stack, 247 Operating system, 10, 427, 470-500 CP/M, 24 history, 9-12 OS/2, 25 UNIX, 470-473, 479,480, 482-489, 494-497 Windows, 473-479, 480-482, 489-493, 497-500 Operating system machine level, 7, 427-505 Operating system macro, 11 Operation, 551, 641 Operation code, 232 Optical disk, 93-102 Orange Book, 97 Orca, 640-642 OS/2, 25

INDEX Osborne-1, 25, Out-of-order execution, 304-309 Overlay, 429

P Packet, 213, 569, 614 Packet processing engine, 572 Packet switching, 569 Page, 430 Page directory, 447 Page fault, 433 Page frame, 432 Page replacement algorithm, 436-438 FIFO, 437 LRU, 436 Page scanner, 604 Page size, 438-439 Page table, 430 Paging, 429-453 demand, 433-436 implementation, 431-433 Parallel computer architecture, 547-654 coprocessors, 567-582 instruction-level parallelism, 549-556 media processor, 576-581 multithreading, 556-562 multicomputer,612-649 multiprocessor, 582-612 network processor, 568-575 on-chip parallelism, 548-567 single-chip multiprocessor, 562-567 taxonomy, 587-589 Parallel input/output, 221 Parallel virtual machine, 633 Parallelism instruction-level, 61-65 processor-level, 65-69 Parent process, 494 Parity bit, 74 Parity flag, 718 Partial address decoding, 225 Partitioned resource sharing, 560 Pascal, Blaise, 14-15

771 Pass, assembler, 523 Passive matrix display, 108 Path, 485 Path length, 272 273-281 Payload, 214 PC (see Personal Computer) PC (see Program Counter) PCI (see Peripheral Component Interconnect) PCI bus, 204-212 PCI express bus, 212-217 PDA (see Personal Digital Assistant) PDP-1, 20 PDP-8, 20 PDP-11, 23 Pentium 4, 189-195 addressing, 371-373 bus, 194-195 data types, 350 instruction formats, 357-358 instructions, 386-389 introduction, 37-42 microarchitecture, 312-317 overview of ISA, 339-341 pinout, 192-194 problems, 413-414 towers of Hanoi, 409-410 virtual memory, 445-450 history, 39-42 photograph, 40 Perfect shuffle, 601 Performance, parallel computer, 643-649 achieveing, 647-649 Amdahl’s law, 645-646 hardware metrics, 643-645 latency hiding, 648-649 software metrics, 645-646 improving, 643-649 Peripheral bus, 566 Peripheral component interconnect, 104, 204-212 signals, 209-211 transactions, 211-213 Perpendicular recording, 83 Personal computer, 34 Personal digital assistant, 26

772

INDEX

Pervasive computing, 26 Phase modulation, 119 Physical address space, 430 Physical layer, 215 Pigment-based ink, 116 Pinout, 173 PIO (see Parallel Input/Output) Pipe, 494 Pipeline, 61 Mic-3, 287 Mic-4, 291 Pentium I, 63-64 Pentium 4, 314-317 seven-stage, 288-292 UltraSPARC, 320-323 Pipeline stage, 62 Pipeline stall, 300 Pipelined data path, 284 Pipelining, 61-63 Pit, 93 Pixel, 109 PLA (see Programmable Logic Array) Plain old telephone service, 120 Playstation 2, 33 Pointer, 362 Pointer and index registers, 708 Poison bit, 311 Polish notation, 365-368 Portable operating system-ix, 471 Position independent code, 538 Positive logic, 145 POSIX (see Portable Operating System-IX) Postfix, 365 POTS (see Plain Old Telephone Service) PPE (see Protocol Processing Engine) Preamble, 83 Precise interrupt, 307 Predication, 418-420 Prefetch buffer, 61 Prefetch cache, 319 Prefetching, 648 Prefix byte, 267, 357, 389 Present/absent bit, 433 Print engine, 114 Printer, 112-117 color, 115-117

Printer (continued) dye-sublimation, 117 laser, 113-114 matrix, 112 monochrome, 112-115 solid ink, 117 Procedure, 397-401 Procedure call instruction, 381-383 Procedure epilog, 400 Procedure prolog, 400 Process, 428 Process creation, 461 Process management UNIX, 494-497 Windows XP, 497-500 Process synchronization, 466-469 Process and thread manager, Windows XP, 477 Processor, 51-59 Processor bandwidth, 63 Processor bus, 566 Processor consistency, 592 Processor-level parallelism, 65-69 Producerconsumer problem, 462 Program, 1 Program counter, 52, 233, 242, 325 Program counter, 704 Program status word, 324, 338, 449 Programmable logic array, 150-152 Programmable Processing Engine, 572 Programmable ROM, 172 Programmed I/O, 383 Progressive scan, 580 PROM (see Programmable ROM) Protocol, 214, 177, 569 Protocol processing engine, 572 Pseudoinstruction, 515-517, 702, 728 PSW (see Program Status Word) Pthread, 495 Public-key cryptography, 581 PVM (see Parallel Virtual Machine)

Q Queueing unit, 290

773

INDEX

R Race condition, 462-466 Radio frequency identification, 30-31 Radix, 681 conversion between, 683-685 Radix number system, 681-683 RAID (see Redundant Array of Inexpensive Disks) RAM (see Random Access Memory) Random access memory, 44, 171-172 dynamic RAM, 171 static RAM, 171 synchronous DRAM, 172 Ranging, 124 Raster scan, 106 RAW dependence, 286 Read-only memory, 44, 171-172 Read/write pointer, 724 Real mode, 339 Recursion, 381 Recursive procedure, 397 Red book, 93 Red Storm, 622-627 Reduced instruction set computer, 58 design principles, 59-61 versus CISC, 58-59 Redundant array of inexpensive disks, 89-93 Reed-Solomon code, 83 Register, 5, 163-164, 337-338, 704 Register addressing, 361-362 Register displacement, 713 Register indirect addressing, 362-363 Register indirect addressing, 712 Register mode, 361 Register renaming, 304-309 Register windows, 343 Register with index, 713 Register with index and displacement, 714 Relative error, 693 Relative path, 486 Release consistency, 593-594 Relocation, dynamic, 536-542 Relocation constant, 533 Relocation problem, 533 Reorder buffer, 315

Replicated worker model, 640 Reserved page, 480 Resource layer, 650 Retirement unit, 317 Reverse Polish notation, 365-368 RFID chip, 30-32 Right justified data, 376 Ring, 616 Ripple carry adder, 155 RISC (see Reduced Instruction Set Computer) RISC design principles, 59-61 RISC vs. CISC, 58-59 ROB (see ReOrder Buffer) ROM (see Read-Only Memory) Root directory, 484 Root hub, USB, 218 Rotational latency, 84 Rounding, 693 Router, 569

S Saturated arithmetic, 553 Scalable processor architecture, 42-43 Scheduler, 316 Scheduling, multicomputer, 635-636 Scoreboard, 305 SCSI (see Small Computer System Interface) SDRAM (see Synchronous DRAM) Seastar, 623 Second-generation computers, 19-21 Second pass, 726 Secondary memory, 81-102 Sector, disk, 83 Security descriptor, 492 Security ID, 491 Security reference monitor, Windows XP, 477 Seek, 84 Segment, 440, 709, 710 Segment override, 725 Segment register group, 709 Segmentation, 439-445 best fit algorithm, 444 first fit algorithm, 444

774

INDEX

Self-modifying, 363 Semantics of memory (see Memory semantics) Semaphore, 466-469 Sequencer, 241 Sequential consistency, 590-591 Serial ATA, 87 Server, 34 Server farm, 36 Session, CD-ROM, 98 Set-associative cache, 297-298 Shard, 629 Shared library, 542 Shared memory (see Multiprocessor) Shared memory, application-level, 636-643 Shared-memory multiprocessor, 582-612 Shell, 473 Shifter, 152-154 Shockley, William, 19 SIB (see Scale, Index, Base) SID (see Security ID) Sign extension, 238 Signed magnitude, 685 Significand, 696 SIMD (see Single Instruction-stream Multiple Data-stream) SIMM (see Single Inline Memory Module) Simple COMA, 612 Simplex, 119 Simulator, 8088, 732-735 Simultaneous multithreading, 558 Single inline memory module, 80 Single instruction stream multiple data stream computer, 66, 587-588 Single large expensive disk, 89 Single-chip multiprocessor, 562-567 Skew, bus, 179 Slave, bus, 177 SLED (see Single Large Expensive Disk) Small computer system interface, 88-89 Small model, 8088, 730 Small outline DIMM, 81 SMP (see Symmetric MultiProcessor) Snooping cache, 190, 595-597 Snoopy cache, 190, 595-597 SO-DIMM (see Small outline DIMM) Socket, 471

Software, 8 communication, 632-635 equivalence with hardware, 8 scheduling, 635-636 Software layer, 216 Solaris, 471 Solid ink printer, 117 Source index, 708 Source language, 507 Source operand, 711 SP (see stack pointer) SPARC (see Scalable PRocessor ARChitecture) Spatial locality, 294 Speculative execution, 309-311 Speculative loads, 420-421 Split cache, 80, 293 Splitter, 121 SR latch, 159-161 SRAM (see Static RAM) SSE (see Streaming SIMD Extensions) Stack, 246-248 Stack addressing, 365-368 Stack frame , 708 Stack pointer, 233, 242, 247, 260, 708 Stage, pipeline, 62 Stale data, 595 Stalled, pipeline, 300 Stalling, 286 Standard error, 484 Standard input, 484 Standard output, 484 Star, 616 State finite state machine, 278 8051, 325 Static RAM, 171 Stibbitz, George, 16 Storage, 69 Store (see memory) Store-and-forward packet switching, 569 Store-to-load, 317 Stream, 472 Streaming SIMD Extensions, 39 Strict consistency, 590 Striping, 90 Strobe, 161

INDEX Structured computer organization, 2 Subroutine, 381, 721 Sun Fire E25K, 606-611 Sun Microsystems, 42-44 Sun UltraSPARC (see UltraSPARC) Supercomputer, 21, 36 Superscalar architecture, 64-65 Superuser, 487 Supervisor call, 11 Switching algebra, 138 Switching network, 600-602 Symbol table, 523, 529-530, 726 Symbolic name, 702 Symmetric key cryptography, 581 Symmetric multiprocessor, 583, 594-602 Synchronous bus, 180-182 Synchronous DRAM, 172 Synchronous message passing, 633 System bus, 176 System call, 11, 427 System interface, 477 System services, Windows XP, 477 System-on-a-chip, 562-567 Systems programmer, 7

T Target, PCI, 207 Target language, 507 Target library, 542 Task bag, 640 TAT-12/13, 28 Taxonomy of parallel computers, 587-589 TCP (see Transmission Control Protocol) Telco, 119 Telecommunications equipment, 117-125 Telephone companie, 119 Template, 639 Temporal locality, 294 Text section, 727 TFT display, 109 Thin film transistor, 109 Third-generation computers, 22-23 Thrashing, 438

775 Thread, 495 Three-bus architecture, 275-276 Threshold sharing, 561 Tightly coupled, 548 Timesharing system, 11 Tiny model, 730 TLB (see Translation Lookaside Buffer) TN (see Twisted Nematic) Token, 565 Top of stack pointer, 233, 242, 260 TOS (see Top Of Stack pointer) Towers of Hanoi, 408-411 Pentium, 409-410 UltraSPARC, 409, 412 Trace BTB, 314 Trace cache, 313 Tracer, 703 8088, 732-735 Transaction layer, 216 Transistor, invention, 19-20 Transistor-transistor logic, 138 Transition, 278 Translation, 2-4 Translation lookaside buffer, 320, 450-452 Miss, 451 Translation storage buffer, 451 Translation table, 452 Translator, 507 Transmission control protocol, 569 header, 570 Transparency, 406 Transparent, 431 Trap, 404, 716 Trap handler, 404 Tree, 616 Tri-state device, 167 TriMedia CPU, 551-556 Triple indirect block, 488 True dependence, 286 Truth table, 138 TSB (see Translation Storage Buffer) TTL (see Transistor-Transistor Logic) Tuning, program, 510 Tuple, 638 Tuple space, 638 Twin page, 638

776

INDEX

Twisted Nematic display, 108 Two’s complement arithmetic, 685 Two-pass translator, 523 TX-0, 20 TX-2, 20

U U pipeline, 63 UART (see Universal Asynchronous Receiver Transmitter) Ubiquitous computing, 26 UDB (see UltraSPARC, data buffer) UHCI (see Universal Host Controller Interface) Ultra port architecture, 198 UltraSPARC Data Buffer II, 198 UltraSPARC I, 43 UltraSPARC III, 196-200 addressing, 373 data buffer, 198 data types, 350-351 history, 42-44 instructions, 389-392 instruction formats, 358-359 microarchitecture, 317-323 overview of ISA, 341-344 virtual memory, 450-452 towers of Hanoi, 409, 411 UMA (see Uniform Memory Access) UMA symmetric multiprocessor, 594-602 Underflow error, 693 UNICODE, 128-131 Unified cache, 80 Uniform memory access, 589 Universal asynchronous receiver transmitter, 221 Universal host controller interface, 22 Universal serial bus, 217-220 USB 2.0, 220 Universal synchronous asynchronous receiver Transmitter, 221 UNIX file I/O, 482-489 introduction, 470-473

UNIX (continued) linking, 542 process management, 494-497 virtual memory, 479-480 UPA (see Ultra Port Architecture) Update strategy, 597 USART (see Universal Synchronous Asynchronous Receiver Transmitter) USB (see Universal Serial Bus) User mode, 335

V V pipeline, 63 Vampire tap, 568 VAX, 57 VCI (see Virtual Component Interconnect) Vector processor, 66-67 Vector register, 67 Very large scale integration, 23 Very long instruction word, 549 Video RAM, 109-110 Virtual 8086 mode, 340 Virtual address space, 430 Virtual circuit, 216 Virtual component interconnect, 567 Virtual cut through, 621 Virtual I/O, 453-460 implementation, 455-460 UNIX, 482-489 Windows XP, 489-493 Virtual machine, 3 Virtual memory, 429-453 compared to caching, 452-453 Pentium 4, 445-450 UltraSPARC III, 450-452 UNIX, 479-480 Windows XP, 480-482 Virtual memory manager, Windows XP, 477 Virtual organization, 649 Virtual register, 246 Virtual topology, 634 Virtuous circle, 28 VIS (see Visual Instruction Set)

777

INDEX Visual instruction set, 43 VLIW (see Very Long Instruction Word) VLSI (see Very Large Scale Integration) Volume table of contents, 98 Von Neumann, John, 18 Von Neumann machine, 18 VTOC (see Volume Table Of Contents)

Wozniak, Steve, 24 Write-after-read dependence, 306 Write-after-write dependence, 307 Write-allocate cache, 298, 597 Write-back cache, 298, 597 Write-deferred cache, 298 Write-through cache, 298, 595

W

X

Wait state, 180 WAN (see Wide-Area Network) WAR dependence (see Write-after-read dependence) Wattel, Evert, 701 WAW dependence (see Write-after-write dependence) Wax printer, 117 Weak consistency, 592-593 WEIZAC, 17 Whirlwind I, 19 Wide-area network, 568 Wilkes, Maurice, 9, 56 Win32 API, 478 Win32 subsystem, 477 Winchester disk, 83 Windows, (see also Windows XP) 25, 473 history, 25 linking, 539-541 Windows 95, 473 Windows 98, 474 Windows NT (see Windows New Technology) Windows New Technology, 474 Windows XP file I/O, 489-493 introduction, 473-479 process management, 497-500 virtual memory, 480-482 Wired-OR, 178 Word, 71, 712 Word instruction, 711 Word register, 711 Working directory, 486 Working set, 433-436

X Windows, 473 Xeon, 40

Y Y2K bug, 36 Yellow Book, 95

Z Zilog Z8000, 57 Zuse, Konrad, 16, 17

ABOUT THE AUTHOR Andrew S. Tanenbaum has an S.B. degree from M.I.T. and a Ph.D. from the University of California at Berkeley. He is currently a Professor of Computer Science at the Vrije Universiteit in Amsterdam, The Netherlands, where he heads the Computer Systems Group. Until stepping down in Jan. 2005, for 12 years he had been Dean of the Advanced School for Computing and Imaging, an interuniversity graduate school doing research on advanced parallel, distributed, and imaging systems. In the past, he has done research on compilers, operating systems, networking, and local-area distributed systems. His current research focuses primarily on computer security, especially in operating systems, networks, and large wide-area distributed systems. Together, all these research projects have led to over 100 refereed papers in journals and conference proceedings and five books. Prof. Tanenbaum has also produced a considerable volume of software. He was the principal architect of the Amsterdam Compiler Kit, a widely-used toolkit for writing portable compilers, as well as of MINIX, a small UNIX clone intended for use in student programming labs. This system provided the inspiration and base on which Linux was developed. Together with his Ph.D. students and programmers, he helped design the Amoeba distributed operating system, a highperformance microkernel-based local-area distributed operating system. After that he was one of the designers of Globe, a wide-area distributed system intended to handle a billion users. This software is now available for free via the Internet. His Ph.D. students have gone on to greater glory after getting their degrees. He is very proud of them. In this respect he resembles a mother hen. Prof. Tanenbaum is a Fellow of the ACM, a Fellow of the the IEEE, and a member of the Royal Netherlands Academy of Arts and Sciences. He is also winner of the 1994 ACM Karl V. Karlstrom Outstanding Educator Award, winner of the 1997 ACM/SIGCSE Award for Outstanding Contributions to Computer Science Education, and winner of the 2002 Texty award for excellence in textbooks. In 2004 he was named as one of the five new Academy Professors by the Royal Academy. He is also listed in Who’s Who in the World. His home page on the World Wide Web can be found at URL http://www.cs.vu.nl/~ast/ .

About the Software on the CD-ROM

Using the CD-ROM To view the contents of this CD, open volume or root level “index.html” with a Web browser. Contents of the CD-ROM •

The 8088 Assembler and Tracer Toolkit This can be run any of the following platforms: Microsoft Windows, Intel/Linux (e.g., little endian) platforms or SPARC/Solaris (e.g., big endian) platforms.

•

Mic-1 MMV Simulation Environment This is written in JavaTM and should run on any Java platform.

Software and Hardware System Requirements Please note the following Minimum System Requirements to run the applications supplied on the accompanying CD-ROM. •

A Pentium II 450-MHz (or faster) processor •

Microsoft® Windows® Server 2003, XP Professional, XP Home Edition, XPMedia Center Edition, XP Tablet PC Edition, 2000 Professional (SP3 or later required for installation), 2000 Server (SP3 or later required for instal lation),Windows ME, Windows 98 (1st or 2nd Editions), or

•

Intel/Linux (e.g., little endian) platforms: RedHat 9.0, or

•

SPARC/Solaris (e.g., big endian) platforms: Solaris 10, or

•

Macintosh; Mac OS X 10.3 (Mic-1 MMV Simulation Environment Only)

•

10 MB of available hard-disk space

•

CD-ROM drive

•

Internet Browser (Internet Explorer 6.0, Firefox 1.0, Mozilla 1.7.x, or Safari 1.2)

•

Internet Connection - required only for downloading additional software, such as updates

•

Screen Resolution 800 x 600 (minimum)

•

Sun's JavaTM JRE 1.4 (Only required if the Mic-1 software is installed)

•

Zip decompression utility (Only required if you will be modifying and recompiling the source files of the Mic-1 software)

Prentice Hall License Agreement and Limited Warranty READ THE FOLLOWING TERMS AND CONDITIONS CAREFULLY BEFORE OPENING THIS SOFTWARE PACKAGE. THIS LEGAL DOCUMENT IS AN AGREEMENT BETWEEN YOU AND PRENTICE-HALL, INC. (THE “COMPANY”). BY OPENING THIS SEALED SOFTWARE PACKAGE, YOU ARE AGREEING TO BE BOUND BY THESE TERMS AND CONDITIONS. IF YOU DO NOT AGREE WITH THESE TERMS AND CONDITIONS, DO NOT OPEN THE SOFTWARE PACKAGE. PROMPTLY RETURN THE UNOPENED SOFTWARE PACKAGE AND ALL ACCOMPANYING ITEMS TO THE PLACE YOU OBTAINED THEM FOR A FULL REFUND OF ANY SUMS YOU HAVE PAID. 1.GRANT OF LICENSE: In consideration of your purchase of this book, and your agreement to abide by the terms and conditions of this Agreement, the Company grants to you a nonexclusive right to use and display the copy of the enclosed software program (hereinafter the “SOFTWARE”) on a single computer (i.e., with a single CPU) at a single location so long as you comply with the terms of this Agreement. The Company reserves all rights not expressly granted to you under this Agreement. 2.OWNERSHIP OF SOFTWARE: You own only the magnetic or physical media (the enclosed media) on which the SOFTWARE is recorded or fixed, but the Company and the software developers retain all the rights, title, and ownership to the SOFTWARE recorded on the original media copy(ies) and all subsequent copies of the SOFTWARE, regardless of the form or media on which the original or other copies may exist. This license is not a sale of the original SOFTWARE or any copy to you. 3.COPY RESTRICTIONS: This SOFTWARE and the accompanying printed materials and user manual (the “Documentation”) are the subject of copyright. The individual programs on the media are copyrighted by the authors of each program. Some of the programs on the media include separate licensing agreements. If you intend to use one of these programs, you must read and follow its accompanying license agreement. You may not copy the Documentation or the SOFTWARE, except that you may make a single copy of the SOFTWARE for backup or archival purposes only. You may be held legally responsible for any copying or copyright infringement which is caused or encouraged by your failure to abide by the terms of this restriction. 4.USE RESTRICTIONS: You may not network the SOFTWARE or otherwise use it on more than one computer or computer terminal at the same time. You may physically transfer the SOFTWARE from one computer to another provided that the SOFTWARE is used on only one computer at a time. You may not distribute copies of the SOFTWARE or Documentation to others. You may not reverse engineer, disassemble, decompile, modify, adapt, translate, or create derivative works based on the SOFTWARE or the Documentation without the prior written consent of the Company. 5. TRANSFER RESTRICTIONS: The enclosed SOFTWARE is licensed only to you and may not be transferred to any one else without the prior written consent of the Company. Any unauthorized transfer of the SOFTWARE shall result in the immediate termination of this Agreement. 6. TERMINATION: This license is effective until terminated. This license will terminate automatically without notice from the Company and become null and void if you fail to comply with any provisions or limitations of this license. Upon termination, you shall destroy the Documentation and all copies of the SOFTWARE. All provisions of this Agreement as to warranties, limitation of liability, remedies or damages, and our ownership rights shall survive termination. 7. MISCELLANEOUS: This Agreement shall be construed in accordance with the laws of the United States of America and the State of New York and shall benefit the Company, its affiliates, and assignees. 8.LIMITED WARRANTY AND DISCLAIMER OF WARRANTY: The Company warrants that the SOFTWARE, when properly used in accordance with the Documentation, will operate in substantial conformity with the description of the SOFTWARE set forth in the Documentation. The Company does not warrant that the SOFTWARE will meet your requirements or that the operation of the SOFTWARE will be uninterrupted or error-free. The Company warrants that the media on which the SOFTWARE is delivered shall be free from defects in materials and workmanship under normal use for a period of thirty (30) days from the date of your purchase. Your only remedy and the Company’s only obligation under these limited warranties is, at the Company’s option, return of the warranted item for a refund of any amounts paid by you or replacement of the item. Any replacement of SOFTWARE or media under the warranties shall not extend the original warranty period. The limited warranty set forth above shall not apply to any SOFTWARE which the Company determines in good faith has been subject to misuse, neglect, improper installation, repair, alteration, or damage by you. EXCEPT FOR THE EXPRESSED WARRANTIES SET FORTH ABOVE, THE COMPANY DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. EXCEPT FOR THE EXPRESS WARRANTY SET FORTH ABOVE, THE COMPANY DOES NOT WARRANT, GUARANTEE, OR MAKE ANY REPRESENTATION REGARDING THE USE OR THE RESULTS OF THE USE OF THE SOFTWARE IN TERMS OF ITS CORRECTNESS, ACCURACY, RELIABILITY, CURRENTNESS, OR OTHERWISE. IN NO EVENT, SHALL THE COMPANY OR ITS EMPLOYEES, AGENTS, SUPPLIERS, OR CONTRACTORS BE LIABLE FOR ANY INCIDENTAL, INDIRECT, SPECIAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF OR IN CONNECTION WITH THE LICENSE GRANTED UNDER THIS AGREEMENT, OR FOR LOSS OF USE, LOSS OF DATA, LOSS OF INCOME OR PROFIT, OR OTHER LOSSES, SUSTAINED AS A RESULT OF INJURY TO ANY PERSON, OR LOSS OF OR DAMAGE TO PROPERTY, OR CLAIMS OF THIRD PARTIES, EVEN IF THE COMPANY OR AN AUTHORIZED REPRESENTATIVE OF THE COMPANY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. IN NO EVENT SHALL LIABILITY OF THE COMPANY FOR DAMAGES WITH RESPECT TO THE SOFTWARE EXCEED THE AMOUNTS ACTUALLY PAID BY YOU, IF ANY, FOR THE SOFTWARE. SOME JURISDICTIONS DO NOT ALLOW THE LIMITATION OF IMPLIED WARRANTIES OR LIABILITY FOR INCIDENTAL, INDIRECT, SPECIAL, OR CONSEQUENTIAL DAMAGES, SO THE ABOVE LIMITATIONS MAY NOT ALWAYS APPLY. THE WARRANTIES IN THIS AGREEMENT GIVE YOU SPECIFIC LEGAL RIGHTS AND YOU MAY ALSO HAVE OTHER RIGHTS WHICH VARY IN ACCORDANCE WITH LOCAL LAW. ACKNOWLEDGMENT YOU ACKNOWLEDGE THAT YOU HAVE READ THIS AGREEMENT, UNDERSTAND IT, AND AGREE TO BE BOUND BY ITS TERMS AND CONDITIONS. YOU ALSO AGREE THAT THIS AGREEMENT IS THE COMPLETE AND EXCLUSIVE STATEMENT OF THE AGREEMENT BETWEEN YOU AND THE COMPANY AND SUPERSEDES ALL PROPOSALS OR PRIOR AGREEMENTS, ORAL, OR WRITTEN, AND ANY OTHER COMMUNICATIONS BETWEEN YOU AND THE COMPANY OR ANY REPRESENTATIVE OF THE COMPANY RELATING TO THE SUBJECT MATTER OF THIS AGREEMENT. Should you have any questions concerning this Agreement or if you wish to contact the Company for any reason, please contact in writing at the address below. Robin Short Prentice Hall PTR One Lake Street Upper Saddle River, New Jersey 07458